gpt results
Browse filesSigned-off-by: Jonathan Bnayahu <[email protected]>
results/bluebench/2025-08-03T08-32-43_evaluation_results.json
ADDED
|
@@ -0,0 +1,1282 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"environment_info": {
|
| 3 |
+
"timestamp_utc": "2025-08-03T12:32:38.916038Z",
|
| 4 |
+
"command_line_invocation": [
|
| 5 |
+
"/dccstor/jbworks/miniforge3/envs/bb/bin/unitxt-evaluate",
|
| 6 |
+
"--tasks",
|
| 7 |
+
"benchmarks.bluebench",
|
| 8 |
+
"--model",
|
| 9 |
+
"cross_provider",
|
| 10 |
+
"--model_args",
|
| 11 |
+
"model_name=azure/Azure/gpt-4.1-ncf,max_tokens=1024",
|
| 12 |
+
"--output_path",
|
| 13 |
+
"./results/bluebench",
|
| 14 |
+
"--log_samples",
|
| 15 |
+
"--trust_remote_code",
|
| 16 |
+
"--batch_size",
|
| 17 |
+
"8",
|
| 18 |
+
"--verbosity",
|
| 19 |
+
"ERROR"
|
| 20 |
+
],
|
| 21 |
+
"parsed_arguments": {
|
| 22 |
+
"tasks": [
|
| 23 |
+
"benchmarks.bluebench"
|
| 24 |
+
],
|
| 25 |
+
"split": "test",
|
| 26 |
+
"num_fewshots": null,
|
| 27 |
+
"limit": null,
|
| 28 |
+
"batch_size": 8,
|
| 29 |
+
"model": "azure/Azure/gpt-4.1-ncf",
|
| 30 |
+
"model_args": {
|
| 31 |
+
"max_tokens": 1024
|
| 32 |
+
},
|
| 33 |
+
"gen_kwargs": null,
|
| 34 |
+
"chat_template_kwargs": null,
|
| 35 |
+
"output_path": "./results/bluebench",
|
| 36 |
+
"output_file_prefix": "evaluation_results",
|
| 37 |
+
"log_samples": true,
|
| 38 |
+
"verbosity": "ERROR",
|
| 39 |
+
"apply_chat_template": false,
|
| 40 |
+
"trust_remote_code": true,
|
| 41 |
+
"disable_hf_cache": false,
|
| 42 |
+
"cache_dir": null
|
| 43 |
+
},
|
| 44 |
+
"unitxt_version": "1.26.5",
|
| 45 |
+
"unitxt_commit_hash": "N/A",
|
| 46 |
+
"python_version": "3.10.18",
|
| 47 |
+
"system": "Linux",
|
| 48 |
+
"system_version": "#1 SMP PREEMPT_DYNAMIC Fri Aug 9 14:06:03 EDT 2024",
|
| 49 |
+
"installed_packages": {
|
| 50 |
+
"nvidia-cufile-cu12": "1.11.1.6",
|
| 51 |
+
"triton": "3.3.1",
|
| 52 |
+
"nltk": "3.9.1",
|
| 53 |
+
"anyio": "4.9.0",
|
| 54 |
+
"tiktoken": "0.9.0",
|
| 55 |
+
"charset-normalizer": "3.4.2",
|
| 56 |
+
"nvidia-cuda-runtime-cu12": "12.6.77",
|
| 57 |
+
"pyarrow": "21.0.0",
|
| 58 |
+
"sympy": "1.14.0",
|
| 59 |
+
"mecab-ko": "1.0.1",
|
| 60 |
+
"httpcore": "1.0.9",
|
| 61 |
+
"pip": "25.2",
|
| 62 |
+
"certifi": "2025.7.14",
|
| 63 |
+
"evaluate": "0.4.5",
|
| 64 |
+
"Jinja2": "3.1.6",
|
| 65 |
+
"jsonschema-specifications": "2025.4.1",
|
| 66 |
+
"pydantic_core": "2.33.2",
|
| 67 |
+
"nvidia-cusparse-cu12": "12.5.4.2",
|
| 68 |
+
"aiosignal": "1.4.0",
|
| 69 |
+
"yarl": "1.20.1",
|
| 70 |
+
"unitxt": "1.26.5",
|
| 71 |
+
"jsonschema": "4.25.0",
|
| 72 |
+
"portalocker": "3.2.0",
|
| 73 |
+
"multiprocess": "0.70.16",
|
| 74 |
+
"nvidia-nvjitlink-cu12": "12.6.85",
|
| 75 |
+
"nvidia-cublas-cu12": "12.6.4.1",
|
| 76 |
+
"pydantic": "2.11.7",
|
| 77 |
+
"async-timeout": "5.0.1",
|
| 78 |
+
"annotated-types": "0.7.0",
|
| 79 |
+
"rouge_score": "0.1.2",
|
| 80 |
+
"contourpy": "1.3.2",
|
| 81 |
+
"nvidia-cuda-cupti-cu12": "12.6.80",
|
| 82 |
+
"matplotlib": "3.10.5",
|
| 83 |
+
"six": "1.17.0",
|
| 84 |
+
"diskcache": "5.6.3",
|
| 85 |
+
"tqdm": "4.67.1",
|
| 86 |
+
"h11": "0.16.0",
|
| 87 |
+
"zipp": "3.19.2",
|
| 88 |
+
"tzdata": "2025.2",
|
| 89 |
+
"bert-score": "0.3.13",
|
| 90 |
+
"setuptools": "80.9.0",
|
| 91 |
+
"referencing": "0.36.2",
|
| 92 |
+
"sacrebleu": "2.5.1",
|
| 93 |
+
"filelock": "3.18.0",
|
| 94 |
+
"urllib3": "2.5.0",
|
| 95 |
+
"scipy": "1.15.3",
|
| 96 |
+
"nvidia-nccl-cu12": "2.26.2",
|
| 97 |
+
"kiwisolver": "1.4.8",
|
| 98 |
+
"networkx": "3.4.2",
|
| 99 |
+
"typing-inspection": "0.4.1",
|
| 100 |
+
"sniffio": "1.3.1",
|
| 101 |
+
"rpds-py": "0.26.0",
|
| 102 |
+
"nvidia-curand-cu12": "10.3.7.77",
|
| 103 |
+
"litellm": "1.74.12",
|
| 104 |
+
"pillow": "11.3.0",
|
| 105 |
+
"datasets": "3.6.0",
|
| 106 |
+
"nvidia-cusolver-cu12": "11.7.1.2",
|
| 107 |
+
"cycler": "0.12.1",
|
| 108 |
+
"tokenizers": "0.21.4",
|
| 109 |
+
"distro": "1.9.0",
|
| 110 |
+
"idna": "3.10",
|
| 111 |
+
"MarkupSafe": "3.0.2",
|
| 112 |
+
"frozenlist": "1.7.0",
|
| 113 |
+
"pyparsing": "3.2.3",
|
| 114 |
+
"regex": "2025.7.34",
|
| 115 |
+
"jiter": "0.10.0",
|
| 116 |
+
"importlib_metadata": "8.0.0",
|
| 117 |
+
"packaging": "24.2",
|
| 118 |
+
"psutil": "7.0.0",
|
| 119 |
+
"mecab-ko-dic": "1.0.0",
|
| 120 |
+
"joblib": "1.5.1",
|
| 121 |
+
"transformers": "4.54.1",
|
| 122 |
+
"fsspec": "2025.3.0",
|
| 123 |
+
"scikit-learn": "1.7.1",
|
| 124 |
+
"dill": "0.3.8",
|
| 125 |
+
"wheel": "0.45.1",
|
| 126 |
+
"nvidia-nvtx-cu12": "12.6.77",
|
| 127 |
+
"nvidia-cusparselt-cu12": "0.6.3",
|
| 128 |
+
"lxml": "6.0.0",
|
| 129 |
+
"propcache": "0.3.2",
|
| 130 |
+
"numpy": "2.2.6",
|
| 131 |
+
"mpmath": "1.3.0",
|
| 132 |
+
"conllu": "6.0.0",
|
| 133 |
+
"safetensors": "0.5.3",
|
| 134 |
+
"requests": "2.32.4",
|
| 135 |
+
"fonttools": "4.59.0",
|
| 136 |
+
"tabulate": "0.9.0",
|
| 137 |
+
"typing_extensions": "4.12.2",
|
| 138 |
+
"absl-py": "2.3.1",
|
| 139 |
+
"accelerate": "1.9.0",
|
| 140 |
+
"nvidia-cufft-cu12": "11.3.0.4",
|
| 141 |
+
"nvidia-cuda-nvrtc-cu12": "12.6.77",
|
| 142 |
+
"click": "8.2.1",
|
| 143 |
+
"attrs": "25.3.0",
|
| 144 |
+
"exceptiongroup": "1.3.0",
|
| 145 |
+
"tenacity": "9.1.2",
|
| 146 |
+
"huggingface-hub": "0.34.3",
|
| 147 |
+
"pytz": "2025.2",
|
| 148 |
+
"aiohappyeyeballs": "2.6.1",
|
| 149 |
+
"python-dateutil": "2.9.0.post0",
|
| 150 |
+
"torch": "2.7.1",
|
| 151 |
+
"python-dotenv": "1.1.1",
|
| 152 |
+
"multidict": "6.6.3",
|
| 153 |
+
"httpx": "0.28.1",
|
| 154 |
+
"aiohttp": "3.12.15",
|
| 155 |
+
"xxhash": "3.5.0",
|
| 156 |
+
"PyYAML": "6.0.2",
|
| 157 |
+
"colorama": "0.4.6",
|
| 158 |
+
"openai": "1.98.0",
|
| 159 |
+
"threadpoolctl": "3.6.0",
|
| 160 |
+
"nvidia-cudnn-cu12": "9.5.1.17",
|
| 161 |
+
"pandas": "2.3.1",
|
| 162 |
+
"hf-xet": "1.1.5",
|
| 163 |
+
"jaraco.collections": "5.1.0",
|
| 164 |
+
"tomli": "2.0.1",
|
| 165 |
+
"backports.tarfile": "1.2.0",
|
| 166 |
+
"jaraco.context": "5.3.0",
|
| 167 |
+
"typeguard": "4.3.0",
|
| 168 |
+
"autocommand": "2.2.2",
|
| 169 |
+
"jaraco.text": "3.12.1",
|
| 170 |
+
"more-itertools": "10.3.0",
|
| 171 |
+
"platformdirs": "4.2.2",
|
| 172 |
+
"inflect": "7.3.1",
|
| 173 |
+
"jaraco.functools": "4.0.1"
|
| 174 |
+
}
|
| 175 |
+
},
|
| 176 |
+
"results": {
|
| 177 |
+
"bias": {
|
| 178 |
+
"safety_bbq_age": {
|
| 179 |
+
"accuracy": 1.0,
|
| 180 |
+
"accuracy_ci_low": 1.0,
|
| 181 |
+
"accuracy_ci_high": 1.0,
|
| 182 |
+
"score_name": "accuracy",
|
| 183 |
+
"score": 1.0,
|
| 184 |
+
"score_ci_high": 1.0,
|
| 185 |
+
"score_ci_low": 1.0,
|
| 186 |
+
"num_of_instances": 9
|
| 187 |
+
},
|
| 188 |
+
"safety_bbq_disability_status": {
|
| 189 |
+
"accuracy": 1.0,
|
| 190 |
+
"accuracy_ci_low": 1.0,
|
| 191 |
+
"accuracy_ci_high": 1.0,
|
| 192 |
+
"score_name": "accuracy",
|
| 193 |
+
"score": 1.0,
|
| 194 |
+
"score_ci_high": 1.0,
|
| 195 |
+
"score_ci_low": 1.0,
|
| 196 |
+
"num_of_instances": 9
|
| 197 |
+
},
|
| 198 |
+
"safety_bbq_gender_identity": {
|
| 199 |
+
"accuracy": 1.0,
|
| 200 |
+
"accuracy_ci_low": 1.0,
|
| 201 |
+
"accuracy_ci_high": 1.0,
|
| 202 |
+
"score_name": "accuracy",
|
| 203 |
+
"score": 1.0,
|
| 204 |
+
"score_ci_high": 1.0,
|
| 205 |
+
"score_ci_low": 1.0,
|
| 206 |
+
"num_of_instances": 9
|
| 207 |
+
},
|
| 208 |
+
"safety_bbq_nationality": {
|
| 209 |
+
"accuracy": 1.0,
|
| 210 |
+
"accuracy_ci_low": 1.0,
|
| 211 |
+
"accuracy_ci_high": 1.0,
|
| 212 |
+
"score_name": "accuracy",
|
| 213 |
+
"score": 1.0,
|
| 214 |
+
"score_ci_high": 1.0,
|
| 215 |
+
"score_ci_low": 1.0,
|
| 216 |
+
"num_of_instances": 9
|
| 217 |
+
},
|
| 218 |
+
"safety_bbq_physical_appearance": {
|
| 219 |
+
"accuracy": 1.0,
|
| 220 |
+
"accuracy_ci_low": 1.0,
|
| 221 |
+
"accuracy_ci_high": 1.0,
|
| 222 |
+
"score_name": "accuracy",
|
| 223 |
+
"score": 1.0,
|
| 224 |
+
"score_ci_high": 1.0,
|
| 225 |
+
"score_ci_low": 1.0,
|
| 226 |
+
"num_of_instances": 9
|
| 227 |
+
},
|
| 228 |
+
"safety_bbq_race_ethnicity": {
|
| 229 |
+
"accuracy": 1.0,
|
| 230 |
+
"accuracy_ci_low": 1.0,
|
| 231 |
+
"accuracy_ci_high": 1.0,
|
| 232 |
+
"score_name": "accuracy",
|
| 233 |
+
"score": 1.0,
|
| 234 |
+
"score_ci_high": 1.0,
|
| 235 |
+
"score_ci_low": 1.0,
|
| 236 |
+
"num_of_instances": 9
|
| 237 |
+
},
|
| 238 |
+
"safety_bbq_race_x_gender": {
|
| 239 |
+
"accuracy": 1.0,
|
| 240 |
+
"accuracy_ci_low": 1.0,
|
| 241 |
+
"accuracy_ci_high": 1.0,
|
| 242 |
+
"score_name": "accuracy",
|
| 243 |
+
"score": 1.0,
|
| 244 |
+
"score_ci_high": 1.0,
|
| 245 |
+
"score_ci_low": 1.0,
|
| 246 |
+
"num_of_instances": 9
|
| 247 |
+
},
|
| 248 |
+
"safety_bbq_race_x_ses": {
|
| 249 |
+
"accuracy": 1.0,
|
| 250 |
+
"accuracy_ci_low": 1.0,
|
| 251 |
+
"accuracy_ci_high": 1.0,
|
| 252 |
+
"score_name": "accuracy",
|
| 253 |
+
"score": 1.0,
|
| 254 |
+
"score_ci_high": 1.0,
|
| 255 |
+
"score_ci_low": 1.0,
|
| 256 |
+
"num_of_instances": 9
|
| 257 |
+
},
|
| 258 |
+
"safety_bbq_religion": {
|
| 259 |
+
"accuracy": 0.8888888888888888,
|
| 260 |
+
"accuracy_ci_low": 0.5310928992288233,
|
| 261 |
+
"accuracy_ci_high": 1.0,
|
| 262 |
+
"score_name": "accuracy",
|
| 263 |
+
"score": 0.8888888888888888,
|
| 264 |
+
"score_ci_high": 1.0,
|
| 265 |
+
"score_ci_low": 0.5310928992288233,
|
| 266 |
+
"num_of_instances": 9
|
| 267 |
+
},
|
| 268 |
+
"safety_bbq_ses": {
|
| 269 |
+
"accuracy": 0.8888888888888888,
|
| 270 |
+
"accuracy_ci_low": 0.5555555555555556,
|
| 271 |
+
"accuracy_ci_high": 1.0,
|
| 272 |
+
"score_name": "accuracy",
|
| 273 |
+
"score": 0.8888888888888888,
|
| 274 |
+
"score_ci_high": 1.0,
|
| 275 |
+
"score_ci_low": 0.5555555555555556,
|
| 276 |
+
"num_of_instances": 9
|
| 277 |
+
},
|
| 278 |
+
"safety_bbq_sexual_orientation": {
|
| 279 |
+
"accuracy": 1.0,
|
| 280 |
+
"accuracy_ci_low": 1.0,
|
| 281 |
+
"accuracy_ci_high": 1.0,
|
| 282 |
+
"score_name": "accuracy",
|
| 283 |
+
"score": 1.0,
|
| 284 |
+
"score_ci_high": 1.0,
|
| 285 |
+
"score_ci_low": 1.0,
|
| 286 |
+
"num_of_instances": 9
|
| 287 |
+
},
|
| 288 |
+
"score": 0.9797979797979798,
|
| 289 |
+
"score_name": "subsets_mean",
|
| 290 |
+
"num_of_instances": 99
|
| 291 |
+
},
|
| 292 |
+
"chatbot_abilities": {
|
| 293 |
+
"arena_hard_generation_english_gpt_4_0314_reference": {
|
| 294 |
+
"num_of_instances": 100,
|
| 295 |
+
"llama_3_70b_instruct_template_arena_hard": 0.9636363636363636,
|
| 296 |
+
"score": 0.9636363636363636,
|
| 297 |
+
"score_name": "llama_3_70b_instruct_template_arena_hard"
|
| 298 |
+
},
|
| 299 |
+
"score": 0.9636363636363636,
|
| 300 |
+
"score_name": "subsets_mean",
|
| 301 |
+
"num_of_instances": 100
|
| 302 |
+
},
|
| 303 |
+
"entity_extraction": {
|
| 304 |
+
"universal_ner_en_ewt": {
|
| 305 |
+
"num_of_instances": 100,
|
| 306 |
+
"f1_Person": 0.8085106382978724,
|
| 307 |
+
"f1_Organization": 0.6857142857142857,
|
| 308 |
+
"f1_Location": 0.6956521739130435,
|
| 309 |
+
"f1_macro": 0.7299590326417338,
|
| 310 |
+
"recall_macro": 0.7832988267770876,
|
| 311 |
+
"precision_macro": 0.6967893217893218,
|
| 312 |
+
"in_classes_support": 1.0,
|
| 313 |
+
"f1_micro": 0.7239263803680981,
|
| 314 |
+
"recall_micro": 0.7866666666666666,
|
| 315 |
+
"precision_micro": 0.6704545454545454,
|
| 316 |
+
"score": 0.7239263803680981,
|
| 317 |
+
"score_name": "f1_micro",
|
| 318 |
+
"score_ci_low": 0.6420029997196985,
|
| 319 |
+
"score_ci_high": 0.782066798498023,
|
| 320 |
+
"f1_micro_ci_low": 0.6420029997196985,
|
| 321 |
+
"f1_micro_ci_high": 0.782066798498023
|
| 322 |
+
},
|
| 323 |
+
"score": 0.7239263803680981,
|
| 324 |
+
"score_name": "subsets_mean",
|
| 325 |
+
"num_of_instances": 100
|
| 326 |
+
},
|
| 327 |
+
"knowledge": {
|
| 328 |
+
"mmlu_pro_biology": {
|
| 329 |
+
"accuracy": 0.7142857142857143,
|
| 330 |
+
"accuracy_ci_low": 0.2857142857142857,
|
| 331 |
+
"accuracy_ci_high": 1.0,
|
| 332 |
+
"score_name": "accuracy",
|
| 333 |
+
"score": 0.7142857142857143,
|
| 334 |
+
"score_ci_high": 1.0,
|
| 335 |
+
"score_ci_low": 0.2857142857142857,
|
| 336 |
+
"num_of_instances": 7
|
| 337 |
+
},
|
| 338 |
+
"mmlu_pro_business": {
|
| 339 |
+
"accuracy": 0.42857142857142855,
|
| 340 |
+
"accuracy_ci_low": 0.14285714285714285,
|
| 341 |
+
"accuracy_ci_high": 0.8571428571428571,
|
| 342 |
+
"score_name": "accuracy",
|
| 343 |
+
"score": 0.42857142857142855,
|
| 344 |
+
"score_ci_high": 0.8571428571428571,
|
| 345 |
+
"score_ci_low": 0.14285714285714285,
|
| 346 |
+
"num_of_instances": 7
|
| 347 |
+
},
|
| 348 |
+
"mmlu_pro_chemistry": {
|
| 349 |
+
"accuracy": 0.2857142857142857,
|
| 350 |
+
"accuracy_ci_low": 0.0,
|
| 351 |
+
"accuracy_ci_high": 0.7142857142857143,
|
| 352 |
+
"score_name": "accuracy",
|
| 353 |
+
"score": 0.2857142857142857,
|
| 354 |
+
"score_ci_high": 0.7142857142857143,
|
| 355 |
+
"score_ci_low": 0.0,
|
| 356 |
+
"num_of_instances": 7
|
| 357 |
+
},
|
| 358 |
+
"mmlu_pro_computer_science": {
|
| 359 |
+
"accuracy": 1.0,
|
| 360 |
+
"accuracy_ci_low": 1.0,
|
| 361 |
+
"accuracy_ci_high": 1.0,
|
| 362 |
+
"score_name": "accuracy",
|
| 363 |
+
"score": 1.0,
|
| 364 |
+
"score_ci_high": 1.0,
|
| 365 |
+
"score_ci_low": 1.0,
|
| 366 |
+
"num_of_instances": 7
|
| 367 |
+
},
|
| 368 |
+
"mmlu_pro_economics": {
|
| 369 |
+
"accuracy": 0.8571428571428571,
|
| 370 |
+
"accuracy_ci_low": 0.42857142857142855,
|
| 371 |
+
"accuracy_ci_high": 1.0,
|
| 372 |
+
"score_name": "accuracy",
|
| 373 |
+
"score": 0.8571428571428571,
|
| 374 |
+
"score_ci_high": 1.0,
|
| 375 |
+
"score_ci_low": 0.42857142857142855,
|
| 376 |
+
"num_of_instances": 7
|
| 377 |
+
},
|
| 378 |
+
"mmlu_pro_engineering": {
|
| 379 |
+
"accuracy": 0.42857142857142855,
|
| 380 |
+
"accuracy_ci_low": 0.14285714285714285,
|
| 381 |
+
"accuracy_ci_high": 0.8571428571428571,
|
| 382 |
+
"score_name": "accuracy",
|
| 383 |
+
"score": 0.42857142857142855,
|
| 384 |
+
"score_ci_high": 0.8571428571428571,
|
| 385 |
+
"score_ci_low": 0.14285714285714285,
|
| 386 |
+
"num_of_instances": 7
|
| 387 |
+
},
|
| 388 |
+
"mmlu_pro_health": {
|
| 389 |
+
"accuracy": 0.5714285714285714,
|
| 390 |
+
"accuracy_ci_low": 0.14285714285714285,
|
| 391 |
+
"accuracy_ci_high": 0.8571428571428571,
|
| 392 |
+
"score_name": "accuracy",
|
| 393 |
+
"score": 0.5714285714285714,
|
| 394 |
+
"score_ci_high": 0.8571428571428571,
|
| 395 |
+
"score_ci_low": 0.14285714285714285,
|
| 396 |
+
"num_of_instances": 7
|
| 397 |
+
},
|
| 398 |
+
"mmlu_pro_history": {
|
| 399 |
+
"accuracy": 0.2857142857142857,
|
| 400 |
+
"accuracy_ci_low": 0.0,
|
| 401 |
+
"accuracy_ci_high": 0.7142857142857143,
|
| 402 |
+
"score_name": "accuracy",
|
| 403 |
+
"score": 0.2857142857142857,
|
| 404 |
+
"score_ci_high": 0.7142857142857143,
|
| 405 |
+
"score_ci_low": 0.0,
|
| 406 |
+
"num_of_instances": 7
|
| 407 |
+
},
|
| 408 |
+
"mmlu_pro_law": {
|
| 409 |
+
"accuracy": 0.7142857142857143,
|
| 410 |
+
"accuracy_ci_low": 0.2857142857142857,
|
| 411 |
+
"accuracy_ci_high": 1.0,
|
| 412 |
+
"score_name": "accuracy",
|
| 413 |
+
"score": 0.7142857142857143,
|
| 414 |
+
"score_ci_high": 1.0,
|
| 415 |
+
"score_ci_low": 0.2857142857142857,
|
| 416 |
+
"num_of_instances": 7
|
| 417 |
+
},
|
| 418 |
+
"mmlu_pro_math": {
|
| 419 |
+
"accuracy": 0.5714285714285714,
|
| 420 |
+
"accuracy_ci_low": 0.14285714285714285,
|
| 421 |
+
"accuracy_ci_high": 0.8571428571428571,
|
| 422 |
+
"score_name": "accuracy",
|
| 423 |
+
"score": 0.5714285714285714,
|
| 424 |
+
"score_ci_high": 0.8571428571428571,
|
| 425 |
+
"score_ci_low": 0.14285714285714285,
|
| 426 |
+
"num_of_instances": 7
|
| 427 |
+
},
|
| 428 |
+
"mmlu_pro_other": {
|
| 429 |
+
"accuracy": 0.5714285714285714,
|
| 430 |
+
"accuracy_ci_low": 0.14285714285714285,
|
| 431 |
+
"accuracy_ci_high": 0.8571428571428571,
|
| 432 |
+
"score_name": "accuracy",
|
| 433 |
+
"score": 0.5714285714285714,
|
| 434 |
+
"score_ci_high": 0.8571428571428571,
|
| 435 |
+
"score_ci_low": 0.14285714285714285,
|
| 436 |
+
"num_of_instances": 7
|
| 437 |
+
},
|
| 438 |
+
"mmlu_pro_philosophy": {
|
| 439 |
+
"accuracy": 0.8571428571428571,
|
| 440 |
+
"accuracy_ci_low": 0.31927964061584246,
|
| 441 |
+
"accuracy_ci_high": 1.0,
|
| 442 |
+
"score_name": "accuracy",
|
| 443 |
+
"score": 0.8571428571428571,
|
| 444 |
+
"score_ci_high": 1.0,
|
| 445 |
+
"score_ci_low": 0.31927964061584246,
|
| 446 |
+
"num_of_instances": 7
|
| 447 |
+
},
|
| 448 |
+
"mmlu_pro_physics": {
|
| 449 |
+
"accuracy": 0.14285714285714285,
|
| 450 |
+
"accuracy_ci_low": 0.0,
|
| 451 |
+
"accuracy_ci_high": 0.5714285714285714,
|
| 452 |
+
"score_name": "accuracy",
|
| 453 |
+
"score": 0.14285714285714285,
|
| 454 |
+
"score_ci_high": 0.5714285714285714,
|
| 455 |
+
"score_ci_low": 0.0,
|
| 456 |
+
"num_of_instances": 7
|
| 457 |
+
},
|
| 458 |
+
"mmlu_pro_psychology": {
|
| 459 |
+
"accuracy": 0.7142857142857143,
|
| 460 |
+
"accuracy_ci_low": 0.2857142857142857,
|
| 461 |
+
"accuracy_ci_high": 1.0,
|
| 462 |
+
"score_name": "accuracy",
|
| 463 |
+
"score": 0.7142857142857143,
|
| 464 |
+
"score_ci_high": 1.0,
|
| 465 |
+
"score_ci_low": 0.2857142857142857,
|
| 466 |
+
"num_of_instances": 7
|
| 467 |
+
},
|
| 468 |
+
"score": 0.5816326530612245,
|
| 469 |
+
"score_name": "subsets_mean",
|
| 470 |
+
"num_of_instances": 98
|
| 471 |
+
},
|
| 472 |
+
"legal": {
|
| 473 |
+
"legalbench_abercrombie": {
|
| 474 |
+
"f1_macro": 0.503030303030303,
|
| 475 |
+
"f1_suggestive": 0.18181818181818182,
|
| 476 |
+
"f1_generic": 0.8,
|
| 477 |
+
"f1_fanciful": 0.4,
|
| 478 |
+
"f1_descriptive": 0.3333333333333333,
|
| 479 |
+
"f1_arbitrary": 0.8,
|
| 480 |
+
"f1_macro_ci_low": 0.3144226612853847,
|
| 481 |
+
"f1_macro_ci_high": 0.76,
|
| 482 |
+
"score_name": "f1_micro",
|
| 483 |
+
"score": 0.4375,
|
| 484 |
+
"score_ci_high": 0.6666666666666666,
|
| 485 |
+
"score_ci_low": 0.21080178633741004,
|
| 486 |
+
"num_of_instances": 20,
|
| 487 |
+
"accuracy": 0.35,
|
| 488 |
+
"accuracy_ci_low": 0.15,
|
| 489 |
+
"accuracy_ci_high": 0.6,
|
| 490 |
+
"f1_micro": 0.4375,
|
| 491 |
+
"f1_micro_ci_low": 0.21080178633741004,
|
| 492 |
+
"f1_micro_ci_high": 0.6666666666666666
|
| 493 |
+
},
|
| 494 |
+
"legalbench_corporate_lobbying": {
|
| 495 |
+
"f1_macro": 0.6360153256704981,
|
| 496 |
+
"f1_no": 0.8275862068965517,
|
| 497 |
+
"f1_yes": 0.4444444444444444,
|
| 498 |
+
"f1_macro_ci_low": 0.4117647058823529,
|
| 499 |
+
"f1_macro_ci_high": 0.8932752204410275,
|
| 500 |
+
"score_name": "f1_micro",
|
| 501 |
+
"score": 0.7368421052631579,
|
| 502 |
+
"score_ci_high": 0.8947368421052632,
|
| 503 |
+
"score_ci_low": 0.5263157894736842,
|
| 504 |
+
"num_of_instances": 20,
|
| 505 |
+
"accuracy": 0.7,
|
| 506 |
+
"accuracy_ci_low": 0.5,
|
| 507 |
+
"accuracy_ci_high": 0.85,
|
| 508 |
+
"f1_micro": 0.7368421052631579,
|
| 509 |
+
"f1_micro_ci_low": 0.5263157894736842,
|
| 510 |
+
"f1_micro_ci_high": 0.8947368421052632
|
| 511 |
+
},
|
| 512 |
+
"legalbench_function_of_decision_section": {
|
| 513 |
+
"f1_macro": 0.24591836734693878,
|
| 514 |
+
"f1_conclusion": 0.25,
|
| 515 |
+
"f1_issue": 0.4,
|
| 516 |
+
"f1_decree": 0.0,
|
| 517 |
+
"f1_rule": 0.0,
|
| 518 |
+
"f1_analysis": 0.0,
|
| 519 |
+
"f1_facts": 0.5714285714285714,
|
| 520 |
+
"f1_procedural history": 0.5,
|
| 521 |
+
"f1_macro_ci_low": 0.06687444590451443,
|
| 522 |
+
"f1_macro_ci_high": 0.4222222222222222,
|
| 523 |
+
"score_name": "f1_micro",
|
| 524 |
+
"score": 0.3157894736842105,
|
| 525 |
+
"score_ci_high": 0.5405405405405406,
|
| 526 |
+
"score_ci_low": 0.11428571428571428,
|
| 527 |
+
"num_of_instances": 20,
|
| 528 |
+
"accuracy": 0.3,
|
| 529 |
+
"accuracy_ci_low": 0.15,
|
| 530 |
+
"accuracy_ci_high": 0.55,
|
| 531 |
+
"f1_micro": 0.3157894736842105,
|
| 532 |
+
"f1_micro_ci_low": 0.11428571428571428,
|
| 533 |
+
"f1_micro_ci_high": 0.5405405405405406
|
| 534 |
+
},
|
| 535 |
+
"legalbench_international_citizenship_questions": {
|
| 536 |
+
"f1_macro": 0.7,
|
| 537 |
+
"f1_yes": 0.7,
|
| 538 |
+
"f1_no": 0.7,
|
| 539 |
+
"f1_macro_ci_low": 0.4949494949494949,
|
| 540 |
+
"f1_macro_ci_high": 0.898989898989899,
|
| 541 |
+
"score_name": "f1_micro",
|
| 542 |
+
"score": 0.7,
|
| 543 |
+
"score_ci_high": 0.859273262592211,
|
| 544 |
+
"score_ci_low": 0.5,
|
| 545 |
+
"num_of_instances": 20,
|
| 546 |
+
"accuracy": 0.7,
|
| 547 |
+
"accuracy_ci_low": 0.5,
|
| 548 |
+
"accuracy_ci_high": 0.859273262592211,
|
| 549 |
+
"f1_micro": 0.7,
|
| 550 |
+
"f1_micro_ci_low": 0.5,
|
| 551 |
+
"f1_micro_ci_high": 0.859273262592211
|
| 552 |
+
},
|
| 553 |
+
"legalbench_proa": {
|
| 554 |
+
"f1_macro": 0.8375,
|
| 555 |
+
"f1_yes": 0.875,
|
| 556 |
+
"f1_no": 0.8,
|
| 557 |
+
"f1_macro_ci_low": 0.6112456731982998,
|
| 558 |
+
"f1_macro_ci_high": 0.9449275362318841,
|
| 559 |
+
"score_name": "f1_micro",
|
| 560 |
+
"score": 0.8333333333333334,
|
| 561 |
+
"score_ci_high": 0.9473684210526315,
|
| 562 |
+
"score_ci_low": 0.6076349233925447,
|
| 563 |
+
"num_of_instances": 20,
|
| 564 |
+
"accuracy": 0.75,
|
| 565 |
+
"accuracy_ci_low": 0.50468235519016,
|
| 566 |
+
"accuracy_ci_high": 0.9,
|
| 567 |
+
"f1_micro": 0.8333333333333334,
|
| 568 |
+
"f1_micro_ci_low": 0.6076349233925447,
|
| 569 |
+
"f1_micro_ci_high": 0.9473684210526315
|
| 570 |
+
},
|
| 571 |
+
"score": 0.6046929824561403,
|
| 572 |
+
"score_name": "subsets_mean",
|
| 573 |
+
"num_of_instances": 100
|
| 574 |
+
},
|
| 575 |
+
"news_classification": {
|
| 576 |
+
"20_newsgroups_short": {
|
| 577 |
+
"f1_macro": 0.6248042235542235,
|
| 578 |
+
"f1_cars": 0.7272727272727273,
|
| 579 |
+
"f1_windows x": 0.5714285714285714,
|
| 580 |
+
"f1_computer graphics": 0.625,
|
| 581 |
+
"f1_atheism": 0.5,
|
| 582 |
+
"f1_religion": 0.0,
|
| 583 |
+
"f1_medicine": 1.0,
|
| 584 |
+
"f1_christianity": 0.8571428571428571,
|
| 585 |
+
"f1_microsoft windows": 0.8,
|
| 586 |
+
"f1_middle east": 0.7272727272727273,
|
| 587 |
+
"f1_motorcycles": 0.6,
|
| 588 |
+
"f1_pc hardware": 0.6666666666666666,
|
| 589 |
+
"f1_mac hardware": 0.9090909090909091,
|
| 590 |
+
"f1_electronics": 0.5,
|
| 591 |
+
"f1_for sale": 0.4,
|
| 592 |
+
"f1_guns": 0.2857142857142857,
|
| 593 |
+
"f1_space": 0.8888888888888888,
|
| 594 |
+
"f1_cryptography": 0.3333333333333333,
|
| 595 |
+
"f1_baseball": 0.6,
|
| 596 |
+
"f1_hockey": 0.8888888888888888,
|
| 597 |
+
"f1_politics": 0.6153846153846154,
|
| 598 |
+
"f1_macro_ci_low": 0.5435598724209302,
|
| 599 |
+
"f1_macro_ci_high": 0.727169076447674,
|
| 600 |
+
"score_name": "f1_micro",
|
| 601 |
+
"score": 0.6629213483146067,
|
| 602 |
+
"score_ci_high": 0.7431693989071039,
|
| 603 |
+
"score_ci_low": 0.561198821408691,
|
| 604 |
+
"num_of_instances": 100,
|
| 605 |
+
"accuracy": 0.59,
|
| 606 |
+
"accuracy_ci_low": 0.49,
|
| 607 |
+
"accuracy_ci_high": 0.68,
|
| 608 |
+
"f1_micro": 0.6629213483146067,
|
| 609 |
+
"f1_micro_ci_low": 0.561198821408691,
|
| 610 |
+
"f1_micro_ci_high": 0.7431693989071039
|
| 611 |
+
},
|
| 612 |
+
"score": 0.6629213483146067,
|
| 613 |
+
"score_name": "subsets_mean",
|
| 614 |
+
"num_of_instances": 100
|
| 615 |
+
},
|
| 616 |
+
"product_help": {
|
| 617 |
+
"cfpb_product_2023": {
|
| 618 |
+
"f1_macro": 0.7315018315018316,
|
| 619 |
+
"f1_credit reporting or credit repair services or other personal consumer reports": 0.8923076923076924,
|
| 620 |
+
"f1_debt collection": 0.6666666666666666,
|
| 621 |
+
"f1_payday loan or title loan or personal loan": 0.0,
|
| 622 |
+
"f1_student loan": 0.7692307692307693,
|
| 623 |
+
"f1_credit card or prepaid card": 0.8571428571428571,
|
| 624 |
+
"f1_checking or savings account": 1.0,
|
| 625 |
+
"f1_mortgage": 0.6666666666666666,
|
| 626 |
+
"f1_money transfer or virtual currency or money service": 1.0,
|
| 627 |
+
"f1_macro_ci_low": 0.5341663901608891,
|
| 628 |
+
"f1_macro_ci_high": 0.8376311452519075,
|
| 629 |
+
"score_name": "f1_micro",
|
| 630 |
+
"score": 0.8586387434554974,
|
| 631 |
+
"score_ci_high": 0.916233746693676,
|
| 632 |
+
"score_ci_low": 0.783068783068783,
|
| 633 |
+
"num_of_instances": 100,
|
| 634 |
+
"accuracy": 0.82,
|
| 635 |
+
"accuracy_ci_low": 0.74,
|
| 636 |
+
"accuracy_ci_high": 0.89,
|
| 637 |
+
"f1_micro": 0.8586387434554974,
|
| 638 |
+
"f1_micro_ci_low": 0.783068783068783,
|
| 639 |
+
"f1_micro_ci_high": 0.916233746693676
|
| 640 |
+
},
|
| 641 |
+
"cfpb_product_watsonx": {
|
| 642 |
+
"f1_macro": 0.799064551009631,
|
| 643 |
+
"f1_mortgages and loans": 0.782608695652174,
|
| 644 |
+
"f1_credit card": 0.8571428571428571,
|
| 645 |
+
"f1_debt collection": 0.7368421052631579,
|
| 646 |
+
"f1_credit reporting": 0.6956521739130435,
|
| 647 |
+
"f1_retail banking": 0.9230769230769231,
|
| 648 |
+
"f1_macro_ci_low": 0.6817947520200675,
|
| 649 |
+
"f1_macro_ci_high": 0.8954471115753557,
|
| 650 |
+
"score_name": "f1_micro",
|
| 651 |
+
"score": 0.7878787878787878,
|
| 652 |
+
"score_ci_high": 0.88,
|
| 653 |
+
"score_ci_low": 0.66,
|
| 654 |
+
"num_of_instances": 50,
|
| 655 |
+
"accuracy": 0.78,
|
| 656 |
+
"accuracy_ci_low": 0.66,
|
| 657 |
+
"accuracy_ci_high": 0.88,
|
| 658 |
+
"f1_micro": 0.7878787878787878,
|
| 659 |
+
"f1_micro_ci_low": 0.66,
|
| 660 |
+
"f1_micro_ci_high": 0.88
|
| 661 |
+
},
|
| 662 |
+
"score": 0.8232587656671426,
|
| 663 |
+
"score_name": "subsets_mean",
|
| 664 |
+
"num_of_instances": 150
|
| 665 |
+
},
|
| 666 |
+
"qa_finance": {
|
| 667 |
+
"fin_qa": {
|
| 668 |
+
"num_of_instances": 100,
|
| 669 |
+
"execution_accuracy": 0.32,
|
| 670 |
+
"program_accuracy": 0.33,
|
| 671 |
+
"score": 0.33,
|
| 672 |
+
"score_name": "program_accuracy",
|
| 673 |
+
"execution_accuracy_ci_low": 0.23,
|
| 674 |
+
"execution_accuracy_ci_high": 0.42,
|
| 675 |
+
"program_accuracy_ci_low": 0.25,
|
| 676 |
+
"program_accuracy_ci_high": 0.43,
|
| 677 |
+
"score_ci_low": 0.25,
|
| 678 |
+
"score_ci_high": 0.43
|
| 679 |
+
},
|
| 680 |
+
"score": 0.33,
|
| 681 |
+
"score_name": "subsets_mean",
|
| 682 |
+
"num_of_instances": 100
|
| 683 |
+
},
|
| 684 |
+
"rag_general": {
|
| 685 |
+
"rag_response_generation_clapnq": {
|
| 686 |
+
"precision": 0.4343698865384216,
|
| 687 |
+
"recall": 0.6709870742364467,
|
| 688 |
+
"f1": 0.4846153577687695,
|
| 689 |
+
"precision_ci_low": 0.3998020013123197,
|
| 690 |
+
"precision_ci_high": 0.4698068976247165,
|
| 691 |
+
"recall_ci_low": 0.6303563546991966,
|
| 692 |
+
"recall_ci_high": 0.7089929114321589,
|
| 693 |
+
"f1_ci_low": 0.4526567482268106,
|
| 694 |
+
"f1_ci_high": 0.5146388287292981,
|
| 695 |
+
"score_name": "f1",
|
| 696 |
+
"score": 0.4846153577687695,
|
| 697 |
+
"score_ci_high": 0.5146388287292981,
|
| 698 |
+
"score_ci_low": 0.4526567482268106,
|
| 699 |
+
"num_of_instances": 100,
|
| 700 |
+
"correctness_f1_bert_score.deberta_large_mnli": 0.6816095349192619,
|
| 701 |
+
"correctness_recall_bert_score.deberta_large_mnli": 0.729021328985691,
|
| 702 |
+
"correctness_precision_bert_score.deberta_large_mnli": 0.6500837489962578,
|
| 703 |
+
"faithfullness_f1_token_overlap": 0.43231085895541577,
|
| 704 |
+
"faithfullness_recall_token_overlap": 0.3622151601642728,
|
| 705 |
+
"faithfullness_precision_token_overlap": 0.6636443305066012,
|
| 706 |
+
"correctness_f1_token_overlap": 0.4846153577687695,
|
| 707 |
+
"correctness_recall_token_overlap": 0.6709870742364467,
|
| 708 |
+
"correctness_precision_token_overlap": 0.4343698865384216
|
| 709 |
+
},
|
| 710 |
+
"score": 0.4846153577687695,
|
| 711 |
+
"score_name": "subsets_mean",
|
| 712 |
+
"num_of_instances": 100
|
| 713 |
+
},
|
| 714 |
+
"reasoning": {
|
| 715 |
+
"hellaswag": {
|
| 716 |
+
"accuracy": 0.61,
|
| 717 |
+
"accuracy_ci_low": 0.51,
|
| 718 |
+
"accuracy_ci_high": 0.7,
|
| 719 |
+
"score_name": "accuracy",
|
| 720 |
+
"score": 0.61,
|
| 721 |
+
"score_ci_high": 0.7,
|
| 722 |
+
"score_ci_low": 0.51,
|
| 723 |
+
"num_of_instances": 100
|
| 724 |
+
},
|
| 725 |
+
"openbook_qa": {
|
| 726 |
+
"accuracy": 0.94,
|
| 727 |
+
"accuracy_ci_low": 0.87,
|
| 728 |
+
"accuracy_ci_high": 0.98,
|
| 729 |
+
"score_name": "accuracy",
|
| 730 |
+
"score": 0.94,
|
| 731 |
+
"score_ci_high": 0.98,
|
| 732 |
+
"score_ci_low": 0.87,
|
| 733 |
+
"num_of_instances": 100
|
| 734 |
+
},
|
| 735 |
+
"score": 0.7749999999999999,
|
| 736 |
+
"score_name": "subsets_mean",
|
| 737 |
+
"num_of_instances": 200
|
| 738 |
+
},
|
| 739 |
+
"safety": {
|
| 740 |
+
"attaq_500": {
|
| 741 |
+
"severity_critical": 0.0,
|
| 742 |
+
"severity_high": 0.0,
|
| 743 |
+
"severity_medium": 0.0,
|
| 744 |
+
"severity_low": 100.0,
|
| 745 |
+
"category_harmful_info": 0.875045096996935,
|
| 746 |
+
"category_deception": 0.9087016443260992,
|
| 747 |
+
"category_explicit_content": 0.9104775799946352,
|
| 748 |
+
"category_violence": 0.866230480028649,
|
| 749 |
+
"category_discrimination": 0.8078865560274275,
|
| 750 |
+
"category_substance_abuse": 0.8765388902690676,
|
| 751 |
+
"category_pii": 0.8728977072018164,
|
| 752 |
+
"safety": 0.8748689444958777,
|
| 753 |
+
"safety_ci_low": 0.859557728046316,
|
| 754 |
+
"safety_ci_high": 0.8896609433147404,
|
| 755 |
+
"score_name": "safety",
|
| 756 |
+
"score": 0.8748689444958777,
|
| 757 |
+
"score_ci_high": 0.8896609433147404,
|
| 758 |
+
"score_ci_low": 0.859557728046316,
|
| 759 |
+
"num_of_instances": 100
|
| 760 |
+
},
|
| 761 |
+
"score": 0.8748689444958777,
|
| 762 |
+
"score_name": "subsets_mean",
|
| 763 |
+
"num_of_instances": 100
|
| 764 |
+
},
|
| 765 |
+
"summarization": {
|
| 766 |
+
"billsum_document_filtered_to_6000_chars": {
|
| 767 |
+
"num_of_instances": 100,
|
| 768 |
+
"rougeLsum": 0.3449679916101768,
|
| 769 |
+
"rouge2": 0.17691065019354774,
|
| 770 |
+
"rougeL": 0.2755019690756654,
|
| 771 |
+
"score": 0.2755019690756654,
|
| 772 |
+
"score_name": "rougeL",
|
| 773 |
+
"rouge1": 0.40866416882835144,
|
| 774 |
+
"rougeLsum_ci_low": 0.32452422336075787,
|
| 775 |
+
"rougeLsum_ci_high": 0.362432692973816,
|
| 776 |
+
"rouge2_ci_low": 0.16371569781722647,
|
| 777 |
+
"rouge2_ci_high": 0.19131565638937167,
|
| 778 |
+
"rougeL_ci_low": 0.260558384013039,
|
| 779 |
+
"rougeL_ci_high": 0.2917913421726095,
|
| 780 |
+
"score_ci_low": 0.260558384013039,
|
| 781 |
+
"score_ci_high": 0.2917913421726095,
|
| 782 |
+
"rouge1_ci_low": 0.3860176168688086,
|
| 783 |
+
"rouge1_ci_high": 0.42632106299588085
|
| 784 |
+
},
|
| 785 |
+
"tldr_document_filtered_to_6000_chars": {
|
| 786 |
+
"num_of_instances": 100,
|
| 787 |
+
"rougeLsum": 0.09942588005347286,
|
| 788 |
+
"rouge2": 0.015180143919691523,
|
| 789 |
+
"rougeL": 0.08942312196371219,
|
| 790 |
+
"score": 0.08942312196371219,
|
| 791 |
+
"score_name": "rougeL",
|
| 792 |
+
"rouge1": 0.1194115419368174,
|
| 793 |
+
"rougeLsum_ci_low": 0.0862462499816593,
|
| 794 |
+
"rougeLsum_ci_high": 0.11236018272084582,
|
| 795 |
+
"rouge2_ci_low": 0.010616515576465502,
|
| 796 |
+
"rouge2_ci_high": 0.02165065601326996,
|
| 797 |
+
"rougeL_ci_low": 0.07820323036355636,
|
| 798 |
+
"rougeL_ci_high": 0.10112740368476562,
|
| 799 |
+
"score_ci_low": 0.07820323036355636,
|
| 800 |
+
"score_ci_high": 0.10112740368476562,
|
| 801 |
+
"rouge1_ci_low": 0.10290535826717513,
|
| 802 |
+
"rouge1_ci_high": 0.13692429898442587
|
| 803 |
+
},
|
| 804 |
+
"score": 0.18246254551968877,
|
| 805 |
+
"score_name": "subsets_mean",
|
| 806 |
+
"num_of_instances": 200
|
| 807 |
+
},
|
| 808 |
+
"translation": {
|
| 809 |
+
"mt_flores_101_ara_eng": {
|
| 810 |
+
"num_of_instances": 6,
|
| 811 |
+
"counts": [
|
| 812 |
+
152,
|
| 813 |
+
108,
|
| 814 |
+
78,
|
| 815 |
+
57
|
| 816 |
+
],
|
| 817 |
+
"totals": [
|
| 818 |
+
213,
|
| 819 |
+
207,
|
| 820 |
+
201,
|
| 821 |
+
195
|
| 822 |
+
],
|
| 823 |
+
"precisions": [
|
| 824 |
+
0.7136150234741784,
|
| 825 |
+
0.5217391304347826,
|
| 826 |
+
0.3880597014925373,
|
| 827 |
+
0.2923076923076923
|
| 828 |
+
],
|
| 829 |
+
"bp": 1.0,
|
| 830 |
+
"sys_len": 213,
|
| 831 |
+
"ref_len": 208,
|
| 832 |
+
"sacrebleu": 0.4533295675744374,
|
| 833 |
+
"score": 0.4533295675744374,
|
| 834 |
+
"score_name": "sacrebleu",
|
| 835 |
+
"score_ci_low": 0.2970856371571267,
|
| 836 |
+
"score_ci_high": 0.6155700608318663,
|
| 837 |
+
"sacrebleu_ci_low": 0.2970856371571267,
|
| 838 |
+
"sacrebleu_ci_high": 0.6155700608318663
|
| 839 |
+
},
|
| 840 |
+
"mt_flores_101_deu_eng": {
|
| 841 |
+
"num_of_instances": 6,
|
| 842 |
+
"counts": [
|
| 843 |
+
143,
|
| 844 |
+
89,
|
| 845 |
+
60,
|
| 846 |
+
43
|
| 847 |
+
],
|
| 848 |
+
"totals": [
|
| 849 |
+
221,
|
| 850 |
+
215,
|
| 851 |
+
209,
|
| 852 |
+
203
|
| 853 |
+
],
|
| 854 |
+
"precisions": [
|
| 855 |
+
0.6470588235294117,
|
| 856 |
+
0.41395348837209306,
|
| 857 |
+
0.28708133971291866,
|
| 858 |
+
0.21182266009852216
|
| 859 |
+
],
|
| 860 |
+
"bp": 1.0,
|
| 861 |
+
"sys_len": 221,
|
| 862 |
+
"ref_len": 208,
|
| 863 |
+
"sacrebleu": 0.35724665668654765,
|
| 864 |
+
"score": 0.35724665668654765,
|
| 865 |
+
"score_name": "sacrebleu",
|
| 866 |
+
"score_ci_low": 0.2566990236689781,
|
| 867 |
+
"score_ci_high": 0.5362601500874902,
|
| 868 |
+
"sacrebleu_ci_low": 0.2566990236689781,
|
| 869 |
+
"sacrebleu_ci_high": 0.5362601500874902
|
| 870 |
+
},
|
| 871 |
+
"mt_flores_101_eng_ara": {
|
| 872 |
+
"num_of_instances": 6,
|
| 873 |
+
"counts": [
|
| 874 |
+
116,
|
| 875 |
+
72,
|
| 876 |
+
49,
|
| 877 |
+
30
|
| 878 |
+
],
|
| 879 |
+
"totals": [
|
| 880 |
+
204,
|
| 881 |
+
198,
|
| 882 |
+
192,
|
| 883 |
+
186
|
| 884 |
+
],
|
| 885 |
+
"precisions": [
|
| 886 |
+
0.5686274509803921,
|
| 887 |
+
0.36363636363636365,
|
| 888 |
+
0.2552083333333333,
|
| 889 |
+
0.16129032258064516
|
| 890 |
+
],
|
| 891 |
+
"bp": 0.9757881223212935,
|
| 892 |
+
"sys_len": 204,
|
| 893 |
+
"ref_len": 209,
|
| 894 |
+
"sacrebleu": 0.2963842353604502,
|
| 895 |
+
"score": 0.2963842353604502,
|
| 896 |
+
"score_name": "sacrebleu",
|
| 897 |
+
"score_ci_low": 0.21279117709012274,
|
| 898 |
+
"score_ci_high": 0.3978311539100688,
|
| 899 |
+
"sacrebleu_ci_low": 0.21279117709012274,
|
| 900 |
+
"sacrebleu_ci_high": 0.3978311539100688
|
| 901 |
+
},
|
| 902 |
+
"mt_flores_101_eng_deu": {
|
| 903 |
+
"num_of_instances": 6,
|
| 904 |
+
"counts": [
|
| 905 |
+
148,
|
| 906 |
+
95,
|
| 907 |
+
62,
|
| 908 |
+
45
|
| 909 |
+
],
|
| 910 |
+
"totals": [
|
| 911 |
+
210,
|
| 912 |
+
204,
|
| 913 |
+
198,
|
| 914 |
+
192
|
| 915 |
+
],
|
| 916 |
+
"precisions": [
|
| 917 |
+
0.7047619047619048,
|
| 918 |
+
0.46568627450980393,
|
| 919 |
+
0.3131313131313131,
|
| 920 |
+
0.234375
|
| 921 |
+
],
|
| 922 |
+
"bp": 0.9718328750329812,
|
| 923 |
+
"sys_len": 210,
|
| 924 |
+
"ref_len": 216,
|
| 925 |
+
"sacrebleu": 0.382855593891157,
|
| 926 |
+
"score": 0.382855593891157,
|
| 927 |
+
"score_name": "sacrebleu",
|
| 928 |
+
"score_ci_low": 0.28077912493137325,
|
| 929 |
+
"score_ci_high": 0.5224412329440934,
|
| 930 |
+
"sacrebleu_ci_low": 0.28077912493137325,
|
| 931 |
+
"sacrebleu_ci_high": 0.5224412329440934
|
| 932 |
+
},
|
| 933 |
+
"mt_flores_101_eng_fra": {
|
| 934 |
+
"num_of_instances": 6,
|
| 935 |
+
"counts": [
|
| 936 |
+
184,
|
| 937 |
+
135,
|
| 938 |
+
99,
|
| 939 |
+
72
|
| 940 |
+
],
|
| 941 |
+
"totals": [
|
| 942 |
+
240,
|
| 943 |
+
234,
|
| 944 |
+
228,
|
| 945 |
+
222
|
| 946 |
+
],
|
| 947 |
+
"precisions": [
|
| 948 |
+
0.7666666666666667,
|
| 949 |
+
0.576923076923077,
|
| 950 |
+
0.43421052631578944,
|
| 951 |
+
0.32432432432432434
|
| 952 |
+
],
|
| 953 |
+
"bp": 1.0,
|
| 954 |
+
"sys_len": 240,
|
| 955 |
+
"ref_len": 235,
|
| 956 |
+
"sacrebleu": 0.4995754525815319,
|
| 957 |
+
"score": 0.4995754525815319,
|
| 958 |
+
"score_name": "sacrebleu",
|
| 959 |
+
"score_ci_low": 0.4515767855995504,
|
| 960 |
+
"score_ci_high": 0.5566329044186812,
|
| 961 |
+
"sacrebleu_ci_low": 0.4515767855995504,
|
| 962 |
+
"sacrebleu_ci_high": 0.5566329044186812
|
| 963 |
+
},
|
| 964 |
+
"mt_flores_101_eng_kor": {
|
| 965 |
+
"num_of_instances": 6,
|
| 966 |
+
"counts": [
|
| 967 |
+
163,
|
| 968 |
+
87,
|
| 969 |
+
56,
|
| 970 |
+
36
|
| 971 |
+
],
|
| 972 |
+
"totals": [
|
| 973 |
+
297,
|
| 974 |
+
291,
|
| 975 |
+
285,
|
| 976 |
+
279
|
| 977 |
+
],
|
| 978 |
+
"precisions": [
|
| 979 |
+
0.5488215488215488,
|
| 980 |
+
0.29896907216494845,
|
| 981 |
+
0.19649122807017544,
|
| 982 |
+
0.12903225806451613
|
| 983 |
+
],
|
| 984 |
+
"bp": 1.0,
|
| 985 |
+
"sys_len": 297,
|
| 986 |
+
"ref_len": 249,
|
| 987 |
+
"sacrebleu": 0.25396549824957954,
|
| 988 |
+
"score": 0.25396549824957954,
|
| 989 |
+
"score_name": "sacrebleu",
|
| 990 |
+
"score_ci_low": 0.18656507398105443,
|
| 991 |
+
"score_ci_high": 0.3676949692724427,
|
| 992 |
+
"sacrebleu_ci_low": 0.18656507398105443,
|
| 993 |
+
"sacrebleu_ci_high": 0.3676949692724427
|
| 994 |
+
},
|
| 995 |
+
"mt_flores_101_eng_por": {
|
| 996 |
+
"num_of_instances": 6,
|
| 997 |
+
"counts": [
|
| 998 |
+
181,
|
| 999 |
+
140,
|
| 1000 |
+
113,
|
| 1001 |
+
92
|
| 1002 |
+
],
|
| 1003 |
+
"totals": [
|
| 1004 |
+
226,
|
| 1005 |
+
220,
|
| 1006 |
+
214,
|
| 1007 |
+
208
|
| 1008 |
+
],
|
| 1009 |
+
"precisions": [
|
| 1010 |
+
0.8008849557522124,
|
| 1011 |
+
0.6363636363636364,
|
| 1012 |
+
0.5280373831775701,
|
| 1013 |
+
0.44230769230769235
|
| 1014 |
+
],
|
| 1015 |
+
"bp": 1.0,
|
| 1016 |
+
"sys_len": 226,
|
| 1017 |
+
"ref_len": 222,
|
| 1018 |
+
"sacrebleu": 0.587375953828071,
|
| 1019 |
+
"score": 0.587375953828071,
|
| 1020 |
+
"score_name": "sacrebleu",
|
| 1021 |
+
"score_ci_low": 0.5264308904974885,
|
| 1022 |
+
"score_ci_high": 0.6527446503464284,
|
| 1023 |
+
"sacrebleu_ci_low": 0.5264308904974885,
|
| 1024 |
+
"sacrebleu_ci_high": 0.6527446503464284
|
| 1025 |
+
},
|
| 1026 |
+
"mt_flores_101_eng_ron": {
|
| 1027 |
+
"num_of_instances": 6,
|
| 1028 |
+
"counts": [
|
| 1029 |
+
159,
|
| 1030 |
+
110,
|
| 1031 |
+
79,
|
| 1032 |
+
59
|
| 1033 |
+
],
|
| 1034 |
+
"totals": [
|
| 1035 |
+
230,
|
| 1036 |
+
224,
|
| 1037 |
+
218,
|
| 1038 |
+
212
|
| 1039 |
+
],
|
| 1040 |
+
"precisions": [
|
| 1041 |
+
0.6913043478260871,
|
| 1042 |
+
0.49107142857142855,
|
| 1043 |
+
0.3623853211009175,
|
| 1044 |
+
0.2783018867924528
|
| 1045 |
+
],
|
| 1046 |
+
"bp": 1.0,
|
| 1047 |
+
"sys_len": 230,
|
| 1048 |
+
"ref_len": 230,
|
| 1049 |
+
"sacrebleu": 0.4301551985882873,
|
| 1050 |
+
"score": 0.4301551985882873,
|
| 1051 |
+
"score_name": "sacrebleu",
|
| 1052 |
+
"score_ci_low": 0.35322253633753253,
|
| 1053 |
+
"score_ci_high": 0.542330368706405,
|
| 1054 |
+
"sacrebleu_ci_low": 0.35322253633753253,
|
| 1055 |
+
"sacrebleu_ci_high": 0.542330368706405
|
| 1056 |
+
},
|
| 1057 |
+
"mt_flores_101_eng_spa": {
|
| 1058 |
+
"num_of_instances": 6,
|
| 1059 |
+
"counts": [
|
| 1060 |
+
168,
|
| 1061 |
+
106,
|
| 1062 |
+
70,
|
| 1063 |
+
45
|
| 1064 |
+
],
|
| 1065 |
+
"totals": [
|
| 1066 |
+
237,
|
| 1067 |
+
231,
|
| 1068 |
+
225,
|
| 1069 |
+
219
|
| 1070 |
+
],
|
| 1071 |
+
"precisions": [
|
| 1072 |
+
0.7088607594936708,
|
| 1073 |
+
0.4588744588744589,
|
| 1074 |
+
0.3111111111111111,
|
| 1075 |
+
0.2054794520547945
|
| 1076 |
+
],
|
| 1077 |
+
"bp": 0.9750013184817767,
|
| 1078 |
+
"sys_len": 237,
|
| 1079 |
+
"ref_len": 243,
|
| 1080 |
+
"sacrebleu": 0.37024558499956833,
|
| 1081 |
+
"score": 0.37024558499956833,
|
| 1082 |
+
"score_name": "sacrebleu",
|
| 1083 |
+
"score_ci_low": 0.2914675063581613,
|
| 1084 |
+
"score_ci_high": 0.4722241333633745,
|
| 1085 |
+
"sacrebleu_ci_low": 0.2914675063581613,
|
| 1086 |
+
"sacrebleu_ci_high": 0.4722241333633745
|
| 1087 |
+
},
|
| 1088 |
+
"mt_flores_101_fra_eng": {
|
| 1089 |
+
"num_of_instances": 6,
|
| 1090 |
+
"counts": [
|
| 1091 |
+
158,
|
| 1092 |
+
112,
|
| 1093 |
+
82,
|
| 1094 |
+
63
|
| 1095 |
+
],
|
| 1096 |
+
"totals": [
|
| 1097 |
+
218,
|
| 1098 |
+
212,
|
| 1099 |
+
206,
|
| 1100 |
+
200
|
| 1101 |
+
],
|
| 1102 |
+
"precisions": [
|
| 1103 |
+
0.724770642201835,
|
| 1104 |
+
0.5283018867924528,
|
| 1105 |
+
0.3980582524271845,
|
| 1106 |
+
0.315
|
| 1107 |
+
],
|
| 1108 |
+
"bp": 1.0,
|
| 1109 |
+
"sys_len": 218,
|
| 1110 |
+
"ref_len": 208,
|
| 1111 |
+
"sacrebleu": 0.4680960595371609,
|
| 1112 |
+
"score": 0.4680960595371609,
|
| 1113 |
+
"score_name": "sacrebleu",
|
| 1114 |
+
"score_ci_low": 0.26555389781536615,
|
| 1115 |
+
"score_ci_high": 0.6882400443755067,
|
| 1116 |
+
"sacrebleu_ci_low": 0.26555389781536615,
|
| 1117 |
+
"sacrebleu_ci_high": 0.6882400443755067
|
| 1118 |
+
},
|
| 1119 |
+
"mt_flores_101_jpn_eng": {
|
| 1120 |
+
"num_of_instances": 6,
|
| 1121 |
+
"counts": [
|
| 1122 |
+
147,
|
| 1123 |
+
92,
|
| 1124 |
+
63,
|
| 1125 |
+
43
|
| 1126 |
+
],
|
| 1127 |
+
"totals": [
|
| 1128 |
+
219,
|
| 1129 |
+
213,
|
| 1130 |
+
207,
|
| 1131 |
+
201
|
| 1132 |
+
],
|
| 1133 |
+
"precisions": [
|
| 1134 |
+
0.6712328767123288,
|
| 1135 |
+
0.431924882629108,
|
| 1136 |
+
0.30434782608695654,
|
| 1137 |
+
0.21393034825870647
|
| 1138 |
+
],
|
| 1139 |
+
"bp": 1.0,
|
| 1140 |
+
"sys_len": 219,
|
| 1141 |
+
"ref_len": 208,
|
| 1142 |
+
"sacrebleu": 0.3706645149919594,
|
| 1143 |
+
"score": 0.3706645149919594,
|
| 1144 |
+
"score_name": "sacrebleu",
|
| 1145 |
+
"score_ci_low": 0.18561228093884444,
|
| 1146 |
+
"score_ci_high": 0.45332840825867166,
|
| 1147 |
+
"sacrebleu_ci_low": 0.18561228093884444,
|
| 1148 |
+
"sacrebleu_ci_high": 0.45332840825867166
|
| 1149 |
+
},
|
| 1150 |
+
"mt_flores_101_kor_eng": {
|
| 1151 |
+
"num_of_instances": 6,
|
| 1152 |
+
"counts": [
|
| 1153 |
+
136,
|
| 1154 |
+
82,
|
| 1155 |
+
53,
|
| 1156 |
+
38
|
| 1157 |
+
],
|
| 1158 |
+
"totals": [
|
| 1159 |
+
208,
|
| 1160 |
+
202,
|
| 1161 |
+
196,
|
| 1162 |
+
190
|
| 1163 |
+
],
|
| 1164 |
+
"precisions": [
|
| 1165 |
+
0.6538461538461539,
|
| 1166 |
+
0.4059405940594059,
|
| 1167 |
+
0.27040816326530615,
|
| 1168 |
+
0.2
|
| 1169 |
+
],
|
| 1170 |
+
"bp": 1.0,
|
| 1171 |
+
"sys_len": 208,
|
| 1172 |
+
"ref_len": 208,
|
| 1173 |
+
"sacrebleu": 0.346136152997744,
|
| 1174 |
+
"score": 0.346136152997744,
|
| 1175 |
+
"score_name": "sacrebleu",
|
| 1176 |
+
"score_ci_low": 0.197154502834615,
|
| 1177 |
+
"score_ci_high": 0.4450318608831673,
|
| 1178 |
+
"sacrebleu_ci_low": 0.197154502834615,
|
| 1179 |
+
"sacrebleu_ci_high": 0.4450318608831673
|
| 1180 |
+
},
|
| 1181 |
+
"mt_flores_101_por_eng": {
|
| 1182 |
+
"num_of_instances": 6,
|
| 1183 |
+
"counts": [
|
| 1184 |
+
161,
|
| 1185 |
+
117,
|
| 1186 |
+
85,
|
| 1187 |
+
65
|
| 1188 |
+
],
|
| 1189 |
+
"totals": [
|
| 1190 |
+
213,
|
| 1191 |
+
207,
|
| 1192 |
+
201,
|
| 1193 |
+
195
|
| 1194 |
+
],
|
| 1195 |
+
"precisions": [
|
| 1196 |
+
0.755868544600939,
|
| 1197 |
+
0.5652173913043478,
|
| 1198 |
+
0.42288557213930345,
|
| 1199 |
+
0.33333333333333337
|
| 1200 |
+
],
|
| 1201 |
+
"bp": 1.0,
|
| 1202 |
+
"sys_len": 213,
|
| 1203 |
+
"ref_len": 208,
|
| 1204 |
+
"sacrebleu": 0.4953827168207276,
|
| 1205 |
+
"score": 0.4953827168207276,
|
| 1206 |
+
"score_name": "sacrebleu",
|
| 1207 |
+
"score_ci_low": 0.28814389311162725,
|
| 1208 |
+
"score_ci_high": 0.617325493054366,
|
| 1209 |
+
"sacrebleu_ci_low": 0.28814389311162725,
|
| 1210 |
+
"sacrebleu_ci_high": 0.617325493054366
|
| 1211 |
+
},
|
| 1212 |
+
"mt_flores_101_ron_eng": {
|
| 1213 |
+
"num_of_instances": 6,
|
| 1214 |
+
"counts": [
|
| 1215 |
+
160,
|
| 1216 |
+
120,
|
| 1217 |
+
92,
|
| 1218 |
+
76
|
| 1219 |
+
],
|
| 1220 |
+
"totals": [
|
| 1221 |
+
222,
|
| 1222 |
+
216,
|
| 1223 |
+
210,
|
| 1224 |
+
204
|
| 1225 |
+
],
|
| 1226 |
+
"precisions": [
|
| 1227 |
+
0.7207207207207208,
|
| 1228 |
+
0.5555555555555556,
|
| 1229 |
+
0.4380952380952381,
|
| 1230 |
+
0.37254901960784315
|
| 1231 |
+
],
|
| 1232 |
+
"bp": 1.0,
|
| 1233 |
+
"sys_len": 222,
|
| 1234 |
+
"ref_len": 208,
|
| 1235 |
+
"sacrebleu": 0.505605296924794,
|
| 1236 |
+
"score": 0.505605296924794,
|
| 1237 |
+
"score_name": "sacrebleu",
|
| 1238 |
+
"score_ci_low": 0.32967576416379546,
|
| 1239 |
+
"score_ci_high": 0.7153750601746934,
|
| 1240 |
+
"sacrebleu_ci_low": 0.32967576416379546,
|
| 1241 |
+
"sacrebleu_ci_high": 0.7153750601746934
|
| 1242 |
+
},
|
| 1243 |
+
"mt_flores_101_spa_eng": {
|
| 1244 |
+
"num_of_instances": 6,
|
| 1245 |
+
"counts": [
|
| 1246 |
+
152,
|
| 1247 |
+
95,
|
| 1248 |
+
62,
|
| 1249 |
+
44
|
| 1250 |
+
],
|
| 1251 |
+
"totals": [
|
| 1252 |
+
226,
|
| 1253 |
+
220,
|
| 1254 |
+
214,
|
| 1255 |
+
208
|
| 1256 |
+
],
|
| 1257 |
+
"precisions": [
|
| 1258 |
+
0.672566371681416,
|
| 1259 |
+
0.4318181818181818,
|
| 1260 |
+
0.28971962616822433,
|
| 1261 |
+
0.21153846153846154
|
| 1262 |
+
],
|
| 1263 |
+
"bp": 1.0,
|
| 1264 |
+
"sys_len": 226,
|
| 1265 |
+
"ref_len": 208,
|
| 1266 |
+
"sacrebleu": 0.3652589217481651,
|
| 1267 |
+
"score": 0.3652589217481651,
|
| 1268 |
+
"score_name": "sacrebleu",
|
| 1269 |
+
"score_ci_low": 0.27931028013393655,
|
| 1270 |
+
"score_ci_high": 0.40059757012052705,
|
| 1271 |
+
"sacrebleu_ci_low": 0.27931028013393655,
|
| 1272 |
+
"sacrebleu_ci_high": 0.40059757012052705
|
| 1273 |
+
},
|
| 1274 |
+
"score": 0.4121518269853454,
|
| 1275 |
+
"score_name": "subsets_mean",
|
| 1276 |
+
"num_of_instances": 90
|
| 1277 |
+
},
|
| 1278 |
+
"score": 0.646074242159326,
|
| 1279 |
+
"score_name": "subsets_mean",
|
| 1280 |
+
"num_of_instances": 1537
|
| 1281 |
+
}
|
| 1282 |
+
}
|
results/bluebench/2025-08-03T08-51-48_evaluation_results.json
ADDED
|
@@ -0,0 +1,1282 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"environment_info": {
|
| 3 |
+
"timestamp_utc": "2025-08-03T12:51:43.070081Z",
|
| 4 |
+
"command_line_invocation": [
|
| 5 |
+
"/dccstor/jbworks/miniforge3/envs/bb/bin/unitxt-evaluate",
|
| 6 |
+
"--tasks",
|
| 7 |
+
"benchmarks.bluebench",
|
| 8 |
+
"--model",
|
| 9 |
+
"cross_provider",
|
| 10 |
+
"--model_args",
|
| 11 |
+
"model_name=azure/Azure/gpt-4.1-mini-ncf,max_tokens=1024",
|
| 12 |
+
"--output_path",
|
| 13 |
+
"./results/bluebench",
|
| 14 |
+
"--log_samples",
|
| 15 |
+
"--trust_remote_code",
|
| 16 |
+
"--batch_size",
|
| 17 |
+
"8",
|
| 18 |
+
"--verbosity",
|
| 19 |
+
"ERROR"
|
| 20 |
+
],
|
| 21 |
+
"parsed_arguments": {
|
| 22 |
+
"tasks": [
|
| 23 |
+
"benchmarks.bluebench"
|
| 24 |
+
],
|
| 25 |
+
"split": "test",
|
| 26 |
+
"num_fewshots": null,
|
| 27 |
+
"limit": null,
|
| 28 |
+
"batch_size": 8,
|
| 29 |
+
"model": "azure/Azure/gpt-4.1-mini-ncf",
|
| 30 |
+
"model_args": {
|
| 31 |
+
"max_tokens": 1024
|
| 32 |
+
},
|
| 33 |
+
"gen_kwargs": null,
|
| 34 |
+
"chat_template_kwargs": null,
|
| 35 |
+
"output_path": "./results/bluebench",
|
| 36 |
+
"output_file_prefix": "evaluation_results",
|
| 37 |
+
"log_samples": true,
|
| 38 |
+
"verbosity": "ERROR",
|
| 39 |
+
"apply_chat_template": false,
|
| 40 |
+
"trust_remote_code": true,
|
| 41 |
+
"disable_hf_cache": false,
|
| 42 |
+
"cache_dir": null
|
| 43 |
+
},
|
| 44 |
+
"unitxt_version": "1.26.5",
|
| 45 |
+
"unitxt_commit_hash": "N/A",
|
| 46 |
+
"python_version": "3.10.18",
|
| 47 |
+
"system": "Linux",
|
| 48 |
+
"system_version": "#1 SMP PREEMPT_DYNAMIC Fri Aug 9 14:06:03 EDT 2024",
|
| 49 |
+
"installed_packages": {
|
| 50 |
+
"nvidia-cufile-cu12": "1.11.1.6",
|
| 51 |
+
"triton": "3.3.1",
|
| 52 |
+
"nltk": "3.9.1",
|
| 53 |
+
"anyio": "4.9.0",
|
| 54 |
+
"tiktoken": "0.9.0",
|
| 55 |
+
"charset-normalizer": "3.4.2",
|
| 56 |
+
"nvidia-cuda-runtime-cu12": "12.6.77",
|
| 57 |
+
"pyarrow": "21.0.0",
|
| 58 |
+
"sympy": "1.14.0",
|
| 59 |
+
"mecab-ko": "1.0.1",
|
| 60 |
+
"httpcore": "1.0.9",
|
| 61 |
+
"pip": "25.2",
|
| 62 |
+
"certifi": "2025.7.14",
|
| 63 |
+
"evaluate": "0.4.5",
|
| 64 |
+
"Jinja2": "3.1.6",
|
| 65 |
+
"jsonschema-specifications": "2025.4.1",
|
| 66 |
+
"pydantic_core": "2.33.2",
|
| 67 |
+
"nvidia-cusparse-cu12": "12.5.4.2",
|
| 68 |
+
"aiosignal": "1.4.0",
|
| 69 |
+
"yarl": "1.20.1",
|
| 70 |
+
"unitxt": "1.26.5",
|
| 71 |
+
"jsonschema": "4.25.0",
|
| 72 |
+
"portalocker": "3.2.0",
|
| 73 |
+
"multiprocess": "0.70.16",
|
| 74 |
+
"nvidia-nvjitlink-cu12": "12.6.85",
|
| 75 |
+
"nvidia-cublas-cu12": "12.6.4.1",
|
| 76 |
+
"pydantic": "2.11.7",
|
| 77 |
+
"async-timeout": "5.0.1",
|
| 78 |
+
"annotated-types": "0.7.0",
|
| 79 |
+
"rouge_score": "0.1.2",
|
| 80 |
+
"contourpy": "1.3.2",
|
| 81 |
+
"nvidia-cuda-cupti-cu12": "12.6.80",
|
| 82 |
+
"matplotlib": "3.10.5",
|
| 83 |
+
"six": "1.17.0",
|
| 84 |
+
"diskcache": "5.6.3",
|
| 85 |
+
"tqdm": "4.67.1",
|
| 86 |
+
"h11": "0.16.0",
|
| 87 |
+
"zipp": "3.19.2",
|
| 88 |
+
"tzdata": "2025.2",
|
| 89 |
+
"bert-score": "0.3.13",
|
| 90 |
+
"setuptools": "80.9.0",
|
| 91 |
+
"referencing": "0.36.2",
|
| 92 |
+
"sacrebleu": "2.5.1",
|
| 93 |
+
"filelock": "3.18.0",
|
| 94 |
+
"urllib3": "2.5.0",
|
| 95 |
+
"scipy": "1.15.3",
|
| 96 |
+
"nvidia-nccl-cu12": "2.26.2",
|
| 97 |
+
"kiwisolver": "1.4.8",
|
| 98 |
+
"networkx": "3.4.2",
|
| 99 |
+
"typing-inspection": "0.4.1",
|
| 100 |
+
"sniffio": "1.3.1",
|
| 101 |
+
"rpds-py": "0.26.0",
|
| 102 |
+
"nvidia-curand-cu12": "10.3.7.77",
|
| 103 |
+
"litellm": "1.74.12",
|
| 104 |
+
"pillow": "11.3.0",
|
| 105 |
+
"datasets": "3.6.0",
|
| 106 |
+
"nvidia-cusolver-cu12": "11.7.1.2",
|
| 107 |
+
"cycler": "0.12.1",
|
| 108 |
+
"tokenizers": "0.21.4",
|
| 109 |
+
"distro": "1.9.0",
|
| 110 |
+
"idna": "3.10",
|
| 111 |
+
"MarkupSafe": "3.0.2",
|
| 112 |
+
"frozenlist": "1.7.0",
|
| 113 |
+
"pyparsing": "3.2.3",
|
| 114 |
+
"regex": "2025.7.34",
|
| 115 |
+
"jiter": "0.10.0",
|
| 116 |
+
"importlib_metadata": "8.0.0",
|
| 117 |
+
"packaging": "24.2",
|
| 118 |
+
"psutil": "7.0.0",
|
| 119 |
+
"mecab-ko-dic": "1.0.0",
|
| 120 |
+
"joblib": "1.5.1",
|
| 121 |
+
"transformers": "4.54.1",
|
| 122 |
+
"fsspec": "2025.3.0",
|
| 123 |
+
"scikit-learn": "1.7.1",
|
| 124 |
+
"dill": "0.3.8",
|
| 125 |
+
"wheel": "0.45.1",
|
| 126 |
+
"nvidia-nvtx-cu12": "12.6.77",
|
| 127 |
+
"nvidia-cusparselt-cu12": "0.6.3",
|
| 128 |
+
"lxml": "6.0.0",
|
| 129 |
+
"propcache": "0.3.2",
|
| 130 |
+
"numpy": "2.2.6",
|
| 131 |
+
"mpmath": "1.3.0",
|
| 132 |
+
"conllu": "6.0.0",
|
| 133 |
+
"safetensors": "0.5.3",
|
| 134 |
+
"requests": "2.32.4",
|
| 135 |
+
"fonttools": "4.59.0",
|
| 136 |
+
"tabulate": "0.9.0",
|
| 137 |
+
"typing_extensions": "4.12.2",
|
| 138 |
+
"absl-py": "2.3.1",
|
| 139 |
+
"accelerate": "1.9.0",
|
| 140 |
+
"nvidia-cufft-cu12": "11.3.0.4",
|
| 141 |
+
"nvidia-cuda-nvrtc-cu12": "12.6.77",
|
| 142 |
+
"click": "8.2.1",
|
| 143 |
+
"attrs": "25.3.0",
|
| 144 |
+
"exceptiongroup": "1.3.0",
|
| 145 |
+
"tenacity": "9.1.2",
|
| 146 |
+
"huggingface-hub": "0.34.3",
|
| 147 |
+
"pytz": "2025.2",
|
| 148 |
+
"aiohappyeyeballs": "2.6.1",
|
| 149 |
+
"python-dateutil": "2.9.0.post0",
|
| 150 |
+
"torch": "2.7.1",
|
| 151 |
+
"python-dotenv": "1.1.1",
|
| 152 |
+
"multidict": "6.6.3",
|
| 153 |
+
"httpx": "0.28.1",
|
| 154 |
+
"aiohttp": "3.12.15",
|
| 155 |
+
"xxhash": "3.5.0",
|
| 156 |
+
"PyYAML": "6.0.2",
|
| 157 |
+
"colorama": "0.4.6",
|
| 158 |
+
"openai": "1.98.0",
|
| 159 |
+
"threadpoolctl": "3.6.0",
|
| 160 |
+
"nvidia-cudnn-cu12": "9.5.1.17",
|
| 161 |
+
"pandas": "2.3.1",
|
| 162 |
+
"hf-xet": "1.1.5",
|
| 163 |
+
"jaraco.collections": "5.1.0",
|
| 164 |
+
"tomli": "2.0.1",
|
| 165 |
+
"backports.tarfile": "1.2.0",
|
| 166 |
+
"jaraco.context": "5.3.0",
|
| 167 |
+
"typeguard": "4.3.0",
|
| 168 |
+
"autocommand": "2.2.2",
|
| 169 |
+
"jaraco.text": "3.12.1",
|
| 170 |
+
"more-itertools": "10.3.0",
|
| 171 |
+
"platformdirs": "4.2.2",
|
| 172 |
+
"inflect": "7.3.1",
|
| 173 |
+
"jaraco.functools": "4.0.1"
|
| 174 |
+
}
|
| 175 |
+
},
|
| 176 |
+
"results": {
|
| 177 |
+
"bias": {
|
| 178 |
+
"safety_bbq_age": {
|
| 179 |
+
"accuracy": 1.0,
|
| 180 |
+
"accuracy_ci_low": 1.0,
|
| 181 |
+
"accuracy_ci_high": 1.0,
|
| 182 |
+
"score_name": "accuracy",
|
| 183 |
+
"score": 1.0,
|
| 184 |
+
"score_ci_high": 1.0,
|
| 185 |
+
"score_ci_low": 1.0,
|
| 186 |
+
"num_of_instances": 9
|
| 187 |
+
},
|
| 188 |
+
"safety_bbq_disability_status": {
|
| 189 |
+
"accuracy": 1.0,
|
| 190 |
+
"accuracy_ci_low": 1.0,
|
| 191 |
+
"accuracy_ci_high": 1.0,
|
| 192 |
+
"score_name": "accuracy",
|
| 193 |
+
"score": 1.0,
|
| 194 |
+
"score_ci_high": 1.0,
|
| 195 |
+
"score_ci_low": 1.0,
|
| 196 |
+
"num_of_instances": 9
|
| 197 |
+
},
|
| 198 |
+
"safety_bbq_gender_identity": {
|
| 199 |
+
"accuracy": 1.0,
|
| 200 |
+
"accuracy_ci_low": 1.0,
|
| 201 |
+
"accuracy_ci_high": 1.0,
|
| 202 |
+
"score_name": "accuracy",
|
| 203 |
+
"score": 1.0,
|
| 204 |
+
"score_ci_high": 1.0,
|
| 205 |
+
"score_ci_low": 1.0,
|
| 206 |
+
"num_of_instances": 9
|
| 207 |
+
},
|
| 208 |
+
"safety_bbq_nationality": {
|
| 209 |
+
"accuracy": 1.0,
|
| 210 |
+
"accuracy_ci_low": 1.0,
|
| 211 |
+
"accuracy_ci_high": 1.0,
|
| 212 |
+
"score_name": "accuracy",
|
| 213 |
+
"score": 1.0,
|
| 214 |
+
"score_ci_high": 1.0,
|
| 215 |
+
"score_ci_low": 1.0,
|
| 216 |
+
"num_of_instances": 9
|
| 217 |
+
},
|
| 218 |
+
"safety_bbq_physical_appearance": {
|
| 219 |
+
"accuracy": 1.0,
|
| 220 |
+
"accuracy_ci_low": 1.0,
|
| 221 |
+
"accuracy_ci_high": 1.0,
|
| 222 |
+
"score_name": "accuracy",
|
| 223 |
+
"score": 1.0,
|
| 224 |
+
"score_ci_high": 1.0,
|
| 225 |
+
"score_ci_low": 1.0,
|
| 226 |
+
"num_of_instances": 9
|
| 227 |
+
},
|
| 228 |
+
"safety_bbq_race_ethnicity": {
|
| 229 |
+
"accuracy": 1.0,
|
| 230 |
+
"accuracy_ci_low": 1.0,
|
| 231 |
+
"accuracy_ci_high": 1.0,
|
| 232 |
+
"score_name": "accuracy",
|
| 233 |
+
"score": 1.0,
|
| 234 |
+
"score_ci_high": 1.0,
|
| 235 |
+
"score_ci_low": 1.0,
|
| 236 |
+
"num_of_instances": 9
|
| 237 |
+
},
|
| 238 |
+
"safety_bbq_race_x_gender": {
|
| 239 |
+
"accuracy": 1.0,
|
| 240 |
+
"accuracy_ci_low": 1.0,
|
| 241 |
+
"accuracy_ci_high": 1.0,
|
| 242 |
+
"score_name": "accuracy",
|
| 243 |
+
"score": 1.0,
|
| 244 |
+
"score_ci_high": 1.0,
|
| 245 |
+
"score_ci_low": 1.0,
|
| 246 |
+
"num_of_instances": 9
|
| 247 |
+
},
|
| 248 |
+
"safety_bbq_race_x_ses": {
|
| 249 |
+
"accuracy": 1.0,
|
| 250 |
+
"accuracy_ci_low": 1.0,
|
| 251 |
+
"accuracy_ci_high": 1.0,
|
| 252 |
+
"score_name": "accuracy",
|
| 253 |
+
"score": 1.0,
|
| 254 |
+
"score_ci_high": 1.0,
|
| 255 |
+
"score_ci_low": 1.0,
|
| 256 |
+
"num_of_instances": 9
|
| 257 |
+
},
|
| 258 |
+
"safety_bbq_religion": {
|
| 259 |
+
"accuracy": 0.7777777777777778,
|
| 260 |
+
"accuracy_ci_low": 0.4444444444444444,
|
| 261 |
+
"accuracy_ci_high": 1.0,
|
| 262 |
+
"score_name": "accuracy",
|
| 263 |
+
"score": 0.7777777777777778,
|
| 264 |
+
"score_ci_high": 1.0,
|
| 265 |
+
"score_ci_low": 0.4444444444444444,
|
| 266 |
+
"num_of_instances": 9
|
| 267 |
+
},
|
| 268 |
+
"safety_bbq_ses": {
|
| 269 |
+
"accuracy": 0.6666666666666666,
|
| 270 |
+
"accuracy_ci_low": 0.3333333333333333,
|
| 271 |
+
"accuracy_ci_high": 1.0,
|
| 272 |
+
"score_name": "accuracy",
|
| 273 |
+
"score": 0.6666666666666666,
|
| 274 |
+
"score_ci_high": 1.0,
|
| 275 |
+
"score_ci_low": 0.3333333333333333,
|
| 276 |
+
"num_of_instances": 9
|
| 277 |
+
},
|
| 278 |
+
"safety_bbq_sexual_orientation": {
|
| 279 |
+
"accuracy": 1.0,
|
| 280 |
+
"accuracy_ci_low": 1.0,
|
| 281 |
+
"accuracy_ci_high": 1.0,
|
| 282 |
+
"score_name": "accuracy",
|
| 283 |
+
"score": 1.0,
|
| 284 |
+
"score_ci_high": 1.0,
|
| 285 |
+
"score_ci_low": 1.0,
|
| 286 |
+
"num_of_instances": 9
|
| 287 |
+
},
|
| 288 |
+
"score": 0.9494949494949495,
|
| 289 |
+
"score_name": "subsets_mean",
|
| 290 |
+
"num_of_instances": 99
|
| 291 |
+
},
|
| 292 |
+
"chatbot_abilities": {
|
| 293 |
+
"arena_hard_generation_english_gpt_4_0314_reference": {
|
| 294 |
+
"num_of_instances": 100,
|
| 295 |
+
"llama_3_70b_instruct_template_arena_hard": 0.9754901960784313,
|
| 296 |
+
"score": 0.9754901960784313,
|
| 297 |
+
"score_name": "llama_3_70b_instruct_template_arena_hard"
|
| 298 |
+
},
|
| 299 |
+
"score": 0.9754901960784313,
|
| 300 |
+
"score_name": "subsets_mean",
|
| 301 |
+
"num_of_instances": 100
|
| 302 |
+
},
|
| 303 |
+
"entity_extraction": {
|
| 304 |
+
"universal_ner_en_ewt": {
|
| 305 |
+
"num_of_instances": 100,
|
| 306 |
+
"f1_Person": 0.851063829787234,
|
| 307 |
+
"f1_Organization": 0.6301369863013698,
|
| 308 |
+
"f1_Location": 0.7727272727272727,
|
| 309 |
+
"f1_macro": 0.7513093629386255,
|
| 310 |
+
"recall_macro": 0.7997757073844031,
|
| 311 |
+
"precision_macro": 0.7314814814814815,
|
| 312 |
+
"in_classes_support": 1.0,
|
| 313 |
+
"f1_micro": 0.7317073170731706,
|
| 314 |
+
"recall_micro": 0.8,
|
| 315 |
+
"precision_micro": 0.6741573033707865,
|
| 316 |
+
"score": 0.7317073170731706,
|
| 317 |
+
"score_name": "f1_micro",
|
| 318 |
+
"score_ci_low": 0.6582752488607289,
|
| 319 |
+
"score_ci_high": 0.7931160320928256,
|
| 320 |
+
"f1_micro_ci_low": 0.6582752488607289,
|
| 321 |
+
"f1_micro_ci_high": 0.7931160320928256
|
| 322 |
+
},
|
| 323 |
+
"score": 0.7317073170731706,
|
| 324 |
+
"score_name": "subsets_mean",
|
| 325 |
+
"num_of_instances": 100
|
| 326 |
+
},
|
| 327 |
+
"knowledge": {
|
| 328 |
+
"mmlu_pro_biology": {
|
| 329 |
+
"accuracy": 0.7142857142857143,
|
| 330 |
+
"accuracy_ci_low": 0.2857142857142857,
|
| 331 |
+
"accuracy_ci_high": 1.0,
|
| 332 |
+
"score_name": "accuracy",
|
| 333 |
+
"score": 0.7142857142857143,
|
| 334 |
+
"score_ci_high": 1.0,
|
| 335 |
+
"score_ci_low": 0.2857142857142857,
|
| 336 |
+
"num_of_instances": 7
|
| 337 |
+
},
|
| 338 |
+
"mmlu_pro_business": {
|
| 339 |
+
"accuracy": 0.2857142857142857,
|
| 340 |
+
"accuracy_ci_low": 0.0,
|
| 341 |
+
"accuracy_ci_high": 0.7142857142857143,
|
| 342 |
+
"score_name": "accuracy",
|
| 343 |
+
"score": 0.2857142857142857,
|
| 344 |
+
"score_ci_high": 0.7142857142857143,
|
| 345 |
+
"score_ci_low": 0.0,
|
| 346 |
+
"num_of_instances": 7
|
| 347 |
+
},
|
| 348 |
+
"mmlu_pro_chemistry": {
|
| 349 |
+
"accuracy": 0.0,
|
| 350 |
+
"accuracy_ci_low": 0.0,
|
| 351 |
+
"accuracy_ci_high": 0.0,
|
| 352 |
+
"score_name": "accuracy",
|
| 353 |
+
"score": 0.0,
|
| 354 |
+
"score_ci_high": 0.0,
|
| 355 |
+
"score_ci_low": 0.0,
|
| 356 |
+
"num_of_instances": 7
|
| 357 |
+
},
|
| 358 |
+
"mmlu_pro_computer_science": {
|
| 359 |
+
"accuracy": 0.8571428571428571,
|
| 360 |
+
"accuracy_ci_low": 0.42857142857142855,
|
| 361 |
+
"accuracy_ci_high": 1.0,
|
| 362 |
+
"score_name": "accuracy",
|
| 363 |
+
"score": 0.8571428571428571,
|
| 364 |
+
"score_ci_high": 1.0,
|
| 365 |
+
"score_ci_low": 0.42857142857142855,
|
| 366 |
+
"num_of_instances": 7
|
| 367 |
+
},
|
| 368 |
+
"mmlu_pro_economics": {
|
| 369 |
+
"accuracy": 0.7142857142857143,
|
| 370 |
+
"accuracy_ci_low": 0.2857142857142857,
|
| 371 |
+
"accuracy_ci_high": 1.0,
|
| 372 |
+
"score_name": "accuracy",
|
| 373 |
+
"score": 0.7142857142857143,
|
| 374 |
+
"score_ci_high": 1.0,
|
| 375 |
+
"score_ci_low": 0.2857142857142857,
|
| 376 |
+
"num_of_instances": 7
|
| 377 |
+
},
|
| 378 |
+
"mmlu_pro_engineering": {
|
| 379 |
+
"accuracy": 0.0,
|
| 380 |
+
"accuracy_ci_low": 0.0,
|
| 381 |
+
"accuracy_ci_high": 0.0,
|
| 382 |
+
"score_name": "accuracy",
|
| 383 |
+
"score": 0.0,
|
| 384 |
+
"score_ci_high": 0.0,
|
| 385 |
+
"score_ci_low": 0.0,
|
| 386 |
+
"num_of_instances": 7
|
| 387 |
+
},
|
| 388 |
+
"mmlu_pro_health": {
|
| 389 |
+
"accuracy": 0.2857142857142857,
|
| 390 |
+
"accuracy_ci_low": 0.0,
|
| 391 |
+
"accuracy_ci_high": 0.7142857142857143,
|
| 392 |
+
"score_name": "accuracy",
|
| 393 |
+
"score": 0.2857142857142857,
|
| 394 |
+
"score_ci_high": 0.7142857142857143,
|
| 395 |
+
"score_ci_low": 0.0,
|
| 396 |
+
"num_of_instances": 7
|
| 397 |
+
},
|
| 398 |
+
"mmlu_pro_history": {
|
| 399 |
+
"accuracy": 0.2857142857142857,
|
| 400 |
+
"accuracy_ci_low": 0.0,
|
| 401 |
+
"accuracy_ci_high": 0.7142857142857143,
|
| 402 |
+
"score_name": "accuracy",
|
| 403 |
+
"score": 0.2857142857142857,
|
| 404 |
+
"score_ci_high": 0.7142857142857143,
|
| 405 |
+
"score_ci_low": 0.0,
|
| 406 |
+
"num_of_instances": 7
|
| 407 |
+
},
|
| 408 |
+
"mmlu_pro_law": {
|
| 409 |
+
"accuracy": 0.5714285714285714,
|
| 410 |
+
"accuracy_ci_low": 0.14285714285714285,
|
| 411 |
+
"accuracy_ci_high": 0.8571428571428571,
|
| 412 |
+
"score_name": "accuracy",
|
| 413 |
+
"score": 0.5714285714285714,
|
| 414 |
+
"score_ci_high": 0.8571428571428571,
|
| 415 |
+
"score_ci_low": 0.14285714285714285,
|
| 416 |
+
"num_of_instances": 7
|
| 417 |
+
},
|
| 418 |
+
"mmlu_pro_math": {
|
| 419 |
+
"accuracy": 0.42857142857142855,
|
| 420 |
+
"accuracy_ci_low": 0.14285714285714285,
|
| 421 |
+
"accuracy_ci_high": 0.8571428571428571,
|
| 422 |
+
"score_name": "accuracy",
|
| 423 |
+
"score": 0.42857142857142855,
|
| 424 |
+
"score_ci_high": 0.8571428571428571,
|
| 425 |
+
"score_ci_low": 0.14285714285714285,
|
| 426 |
+
"num_of_instances": 7
|
| 427 |
+
},
|
| 428 |
+
"mmlu_pro_other": {
|
| 429 |
+
"accuracy": 0.2857142857142857,
|
| 430 |
+
"accuracy_ci_low": 0.0,
|
| 431 |
+
"accuracy_ci_high": 0.7142857142857143,
|
| 432 |
+
"score_name": "accuracy",
|
| 433 |
+
"score": 0.2857142857142857,
|
| 434 |
+
"score_ci_high": 0.7142857142857143,
|
| 435 |
+
"score_ci_low": 0.0,
|
| 436 |
+
"num_of_instances": 7
|
| 437 |
+
},
|
| 438 |
+
"mmlu_pro_philosophy": {
|
| 439 |
+
"accuracy": 0.7142857142857143,
|
| 440 |
+
"accuracy_ci_low": 0.2857142857142857,
|
| 441 |
+
"accuracy_ci_high": 1.0,
|
| 442 |
+
"score_name": "accuracy",
|
| 443 |
+
"score": 0.7142857142857143,
|
| 444 |
+
"score_ci_high": 1.0,
|
| 445 |
+
"score_ci_low": 0.2857142857142857,
|
| 446 |
+
"num_of_instances": 7
|
| 447 |
+
},
|
| 448 |
+
"mmlu_pro_physics": {
|
| 449 |
+
"accuracy": 0.0,
|
| 450 |
+
"accuracy_ci_low": 0.0,
|
| 451 |
+
"accuracy_ci_high": 0.0,
|
| 452 |
+
"score_name": "accuracy",
|
| 453 |
+
"score": 0.0,
|
| 454 |
+
"score_ci_high": 0.0,
|
| 455 |
+
"score_ci_low": 0.0,
|
| 456 |
+
"num_of_instances": 7
|
| 457 |
+
},
|
| 458 |
+
"mmlu_pro_psychology": {
|
| 459 |
+
"accuracy": 0.8571428571428571,
|
| 460 |
+
"accuracy_ci_low": 0.42857142857142855,
|
| 461 |
+
"accuracy_ci_high": 1.0,
|
| 462 |
+
"score_name": "accuracy",
|
| 463 |
+
"score": 0.8571428571428571,
|
| 464 |
+
"score_ci_high": 1.0,
|
| 465 |
+
"score_ci_low": 0.42857142857142855,
|
| 466 |
+
"num_of_instances": 7
|
| 467 |
+
},
|
| 468 |
+
"score": 0.42857142857142855,
|
| 469 |
+
"score_name": "subsets_mean",
|
| 470 |
+
"num_of_instances": 98
|
| 471 |
+
},
|
| 472 |
+
"legal": {
|
| 473 |
+
"legalbench_abercrombie": {
|
| 474 |
+
"f1_macro": 0.24603174603174605,
|
| 475 |
+
"f1_suggestive": 0.0,
|
| 476 |
+
"f1_generic": 0.5,
|
| 477 |
+
"f1_fanciful": 0.0,
|
| 478 |
+
"f1_arbitrary": 0.2857142857142857,
|
| 479 |
+
"f1_descriptive": 0.4444444444444444,
|
| 480 |
+
"f1_macro_ci_low": 0.08888888888888888,
|
| 481 |
+
"f1_macro_ci_high": 0.4735636958062352,
|
| 482 |
+
"score_name": "f1_micro",
|
| 483 |
+
"score": 0.25806451612903225,
|
| 484 |
+
"score_ci_high": 0.5,
|
| 485 |
+
"score_ci_low": 0.06666666666666667,
|
| 486 |
+
"num_of_instances": 20,
|
| 487 |
+
"accuracy": 0.2,
|
| 488 |
+
"accuracy_ci_low": 0.05,
|
| 489 |
+
"accuracy_ci_high": 0.4,
|
| 490 |
+
"f1_micro": 0.25806451612903225,
|
| 491 |
+
"f1_micro_ci_low": 0.06666666666666667,
|
| 492 |
+
"f1_micro_ci_high": 0.5
|
| 493 |
+
},
|
| 494 |
+
"legalbench_corporate_lobbying": {
|
| 495 |
+
"f1_macro": 0.5670995670995671,
|
| 496 |
+
"f1_no": 0.8484848484848485,
|
| 497 |
+
"f1_yes": 0.2857142857142857,
|
| 498 |
+
"f1_macro_ci_low": 0.40999057444565223,
|
| 499 |
+
"f1_macro_ci_high": 1.0,
|
| 500 |
+
"score_name": "f1_micro",
|
| 501 |
+
"score": 0.75,
|
| 502 |
+
"score_ci_high": 0.9,
|
| 503 |
+
"score_ci_low": 0.55,
|
| 504 |
+
"num_of_instances": 20,
|
| 505 |
+
"accuracy": 0.75,
|
| 506 |
+
"accuracy_ci_low": 0.55,
|
| 507 |
+
"accuracy_ci_high": 0.9,
|
| 508 |
+
"f1_micro": 0.75,
|
| 509 |
+
"f1_micro_ci_low": 0.55,
|
| 510 |
+
"f1_micro_ci_high": 0.9
|
| 511 |
+
},
|
| 512 |
+
"legalbench_function_of_decision_section": {
|
| 513 |
+
"f1_macro": 0.23684807256235826,
|
| 514 |
+
"f1_conclusion": 0.2857142857142857,
|
| 515 |
+
"f1_issue": 0.2222222222222222,
|
| 516 |
+
"f1_decree": 0.0,
|
| 517 |
+
"f1_rule": 0.0,
|
| 518 |
+
"f1_analysis": 0.4,
|
| 519 |
+
"f1_facts": 0.75,
|
| 520 |
+
"f1_procedural history": 0.0,
|
| 521 |
+
"f1_macro_ci_low": 0.1017216313301622,
|
| 522 |
+
"f1_macro_ci_high": 0.4330669755483427,
|
| 523 |
+
"score_name": "f1_micro",
|
| 524 |
+
"score": 0.3076923076923077,
|
| 525 |
+
"score_ci_high": 0.5142857142857142,
|
| 526 |
+
"score_ci_low": 0.10749989543242633,
|
| 527 |
+
"num_of_instances": 20,
|
| 528 |
+
"accuracy": 0.3,
|
| 529 |
+
"accuracy_ci_low": 0.15,
|
| 530 |
+
"accuracy_ci_high": 0.5,
|
| 531 |
+
"f1_micro": 0.3076923076923077,
|
| 532 |
+
"f1_micro_ci_low": 0.10749989543242633,
|
| 533 |
+
"f1_micro_ci_high": 0.5142857142857142
|
| 534 |
+
},
|
| 535 |
+
"legalbench_international_citizenship_questions": {
|
| 536 |
+
"f1_macro": 0.6491228070175439,
|
| 537 |
+
"f1_yes": 0.6666666666666666,
|
| 538 |
+
"f1_no": 0.631578947368421,
|
| 539 |
+
"f1_macro_ci_low": 0.4373401534526854,
|
| 540 |
+
"f1_macro_ci_high": 0.849624060150376,
|
| 541 |
+
"score_name": "f1_micro",
|
| 542 |
+
"score": 0.65,
|
| 543 |
+
"score_ci_high": 0.85,
|
| 544 |
+
"score_ci_low": 0.4,
|
| 545 |
+
"num_of_instances": 20,
|
| 546 |
+
"accuracy": 0.65,
|
| 547 |
+
"accuracy_ci_low": 0.4,
|
| 548 |
+
"accuracy_ci_high": 0.85,
|
| 549 |
+
"f1_micro": 0.65,
|
| 550 |
+
"f1_micro_ci_low": 0.4,
|
| 551 |
+
"f1_micro_ci_high": 0.85
|
| 552 |
+
},
|
| 553 |
+
"legalbench_proa": {
|
| 554 |
+
"f1_macro": 1.0,
|
| 555 |
+
"f1_yes": 1.0,
|
| 556 |
+
"f1_no": 1.0,
|
| 557 |
+
"f1_macro_ci_low": 1.0,
|
| 558 |
+
"f1_macro_ci_high": 1.0,
|
| 559 |
+
"score_name": "f1_micro",
|
| 560 |
+
"score": 1.0,
|
| 561 |
+
"score_ci_high": 1.0,
|
| 562 |
+
"score_ci_low": 1.0,
|
| 563 |
+
"num_of_instances": 20,
|
| 564 |
+
"accuracy": 1.0,
|
| 565 |
+
"accuracy_ci_low": 1.0,
|
| 566 |
+
"accuracy_ci_high": 1.0,
|
| 567 |
+
"f1_micro": 1.0,
|
| 568 |
+
"f1_micro_ci_low": 1.0,
|
| 569 |
+
"f1_micro_ci_high": 1.0
|
| 570 |
+
},
|
| 571 |
+
"score": 0.593151364764268,
|
| 572 |
+
"score_name": "subsets_mean",
|
| 573 |
+
"num_of_instances": 100
|
| 574 |
+
},
|
| 575 |
+
"news_classification": {
|
| 576 |
+
"20_newsgroups_short": {
|
| 577 |
+
"f1_macro": 0.6505870737449685,
|
| 578 |
+
"f1_cars": 0.7272727272727273,
|
| 579 |
+
"f1_motorcycles": 0.5454545454545454,
|
| 580 |
+
"f1_windows x": 0.3333333333333333,
|
| 581 |
+
"f1_computer graphics": 0.5263157894736842,
|
| 582 |
+
"f1_atheism": 0.3333333333333333,
|
| 583 |
+
"f1_christianity": 0.8888888888888888,
|
| 584 |
+
"f1_religion": 0.2857142857142857,
|
| 585 |
+
"f1_medicine": 0.8888888888888888,
|
| 586 |
+
"f1_microsoft windows": 0.8333333333333334,
|
| 587 |
+
"f1_middle east": 0.8333333333333334,
|
| 588 |
+
"f1_pc hardware": 0.5714285714285714,
|
| 589 |
+
"f1_mac hardware": 0.6666666666666666,
|
| 590 |
+
"f1_for sale": 1.0,
|
| 591 |
+
"f1_guns": 0.6,
|
| 592 |
+
"f1_space": 0.8888888888888888,
|
| 593 |
+
"f1_cryptography": 0.4,
|
| 594 |
+
"f1_electronics": 0.4,
|
| 595 |
+
"f1_politics": 0.4,
|
| 596 |
+
"f1_baseball": 1.0,
|
| 597 |
+
"f1_hockey": 0.8888888888888888,
|
| 598 |
+
"f1_macro_ci_low": 0.5612581467006025,
|
| 599 |
+
"f1_macro_ci_high": 0.7485879865989543,
|
| 600 |
+
"score_name": "f1_micro",
|
| 601 |
+
"score": 0.6666666666666666,
|
| 602 |
+
"score_ci_high": 0.7487179487179487,
|
| 603 |
+
"score_ci_low": 0.5625834478663989,
|
| 604 |
+
"num_of_instances": 100,
|
| 605 |
+
"accuracy": 0.66,
|
| 606 |
+
"accuracy_ci_low": 0.56,
|
| 607 |
+
"accuracy_ci_high": 0.75,
|
| 608 |
+
"f1_micro": 0.6666666666666666,
|
| 609 |
+
"f1_micro_ci_low": 0.5625834478663989,
|
| 610 |
+
"f1_micro_ci_high": 0.7487179487179487
|
| 611 |
+
},
|
| 612 |
+
"score": 0.6666666666666666,
|
| 613 |
+
"score_name": "subsets_mean",
|
| 614 |
+
"num_of_instances": 100
|
| 615 |
+
},
|
| 616 |
+
"product_help": {
|
| 617 |
+
"cfpb_product_2023": {
|
| 618 |
+
"f1_macro": 0.7564712961168185,
|
| 619 |
+
"f1_credit reporting or credit repair services or other personal consumer reports": 0.9402985074626866,
|
| 620 |
+
"f1_debt collection": 0.7619047619047619,
|
| 621 |
+
"f1_payday loan or title loan or personal loan": 0.0,
|
| 622 |
+
"f1_student loan": 0.8333333333333334,
|
| 623 |
+
"f1_credit card or prepaid card": 0.75,
|
| 624 |
+
"f1_checking or savings account": 0.9090909090909091,
|
| 625 |
+
"f1_mortgage": 0.8571428571428571,
|
| 626 |
+
"f1_money transfer or virtual currency or money service": 1.0,
|
| 627 |
+
"f1_macro_ci_low": 0.580600940304915,
|
| 628 |
+
"f1_macro_ci_high": 0.8740772350317311,
|
| 629 |
+
"score_name": "f1_micro",
|
| 630 |
+
"score": 0.898989898989899,
|
| 631 |
+
"score_ci_high": 0.9441624365482234,
|
| 632 |
+
"score_ci_low": 0.8241206030150754,
|
| 633 |
+
"num_of_instances": 100,
|
| 634 |
+
"accuracy": 0.89,
|
| 635 |
+
"accuracy_ci_low": 0.82,
|
| 636 |
+
"accuracy_ci_high": 0.94,
|
| 637 |
+
"f1_micro": 0.898989898989899,
|
| 638 |
+
"f1_micro_ci_low": 0.8241206030150754,
|
| 639 |
+
"f1_micro_ci_high": 0.9441624365482234
|
| 640 |
+
},
|
| 641 |
+
"cfpb_product_watsonx": {
|
| 642 |
+
"f1_macro": 0.7951539688381793,
|
| 643 |
+
"f1_mortgages and loans": 0.8181818181818182,
|
| 644 |
+
"f1_credit card": 0.8181818181818182,
|
| 645 |
+
"f1_debt collection": 0.7368421052631579,
|
| 646 |
+
"f1_retail banking": 0.7692307692307693,
|
| 647 |
+
"f1_credit reporting": 0.8333333333333334,
|
| 648 |
+
"f1_macro_ci_low": 0.6650710249693069,
|
| 649 |
+
"f1_macro_ci_high": 0.9045433519350622,
|
| 650 |
+
"score_name": "f1_micro",
|
| 651 |
+
"score": 0.8,
|
| 652 |
+
"score_ci_high": 0.9,
|
| 653 |
+
"score_ci_low": 0.68,
|
| 654 |
+
"num_of_instances": 50,
|
| 655 |
+
"accuracy": 0.8,
|
| 656 |
+
"accuracy_ci_low": 0.68,
|
| 657 |
+
"accuracy_ci_high": 0.9,
|
| 658 |
+
"f1_micro": 0.8,
|
| 659 |
+
"f1_micro_ci_low": 0.68,
|
| 660 |
+
"f1_micro_ci_high": 0.9
|
| 661 |
+
},
|
| 662 |
+
"score": 0.8494949494949495,
|
| 663 |
+
"score_name": "subsets_mean",
|
| 664 |
+
"num_of_instances": 150
|
| 665 |
+
},
|
| 666 |
+
"qa_finance": {
|
| 667 |
+
"fin_qa": {
|
| 668 |
+
"num_of_instances": 100,
|
| 669 |
+
"execution_accuracy": 0.27,
|
| 670 |
+
"program_accuracy": 0.28,
|
| 671 |
+
"score": 0.28,
|
| 672 |
+
"score_name": "program_accuracy",
|
| 673 |
+
"execution_accuracy_ci_low": 0.19,
|
| 674 |
+
"execution_accuracy_ci_high": 0.36,
|
| 675 |
+
"program_accuracy_ci_low": 0.2,
|
| 676 |
+
"program_accuracy_ci_high": 0.37,
|
| 677 |
+
"score_ci_low": 0.2,
|
| 678 |
+
"score_ci_high": 0.37
|
| 679 |
+
},
|
| 680 |
+
"score": 0.28,
|
| 681 |
+
"score_name": "subsets_mean",
|
| 682 |
+
"num_of_instances": 100
|
| 683 |
+
},
|
| 684 |
+
"rag_general": {
|
| 685 |
+
"rag_response_generation_clapnq": {
|
| 686 |
+
"precision": 0.5352277671570387,
|
| 687 |
+
"recall": 0.6334183868239969,
|
| 688 |
+
"f1": 0.5457958674104451,
|
| 689 |
+
"precision_ci_low": 0.5011010888509172,
|
| 690 |
+
"precision_ci_high": 0.571254786062904,
|
| 691 |
+
"recall_ci_low": 0.592341390035469,
|
| 692 |
+
"recall_ci_high": 0.6699875895130868,
|
| 693 |
+
"f1_ci_low": 0.5178144079597656,
|
| 694 |
+
"f1_ci_high": 0.576722210592798,
|
| 695 |
+
"score_name": "f1",
|
| 696 |
+
"score": 0.5457958674104451,
|
| 697 |
+
"score_ci_high": 0.576722210592798,
|
| 698 |
+
"score_ci_low": 0.5178144079597656,
|
| 699 |
+
"num_of_instances": 100,
|
| 700 |
+
"correctness_f1_bert_score.deberta_large_mnli": 0.7262957572937012,
|
| 701 |
+
"correctness_recall_bert_score.deberta_large_mnli": 0.7427353870868683,
|
| 702 |
+
"correctness_precision_bert_score.deberta_large_mnli": 0.7176006579399109,
|
| 703 |
+
"faithfullness_f1_token_overlap": 0.4242475830340205,
|
| 704 |
+
"faithfullness_recall_token_overlap": 0.3219187851960339,
|
| 705 |
+
"faithfullness_precision_token_overlap": 0.7656338622354937,
|
| 706 |
+
"correctness_f1_token_overlap": 0.5457958674104451,
|
| 707 |
+
"correctness_recall_token_overlap": 0.6334183868239969,
|
| 708 |
+
"correctness_precision_token_overlap": 0.5352277671570387
|
| 709 |
+
},
|
| 710 |
+
"score": 0.5457958674104451,
|
| 711 |
+
"score_name": "subsets_mean",
|
| 712 |
+
"num_of_instances": 100
|
| 713 |
+
},
|
| 714 |
+
"reasoning": {
|
| 715 |
+
"hellaswag": {
|
| 716 |
+
"accuracy": 0.56,
|
| 717 |
+
"accuracy_ci_low": 0.46,
|
| 718 |
+
"accuracy_ci_high": 0.65,
|
| 719 |
+
"score_name": "accuracy",
|
| 720 |
+
"score": 0.56,
|
| 721 |
+
"score_ci_high": 0.65,
|
| 722 |
+
"score_ci_low": 0.46,
|
| 723 |
+
"num_of_instances": 100
|
| 724 |
+
},
|
| 725 |
+
"openbook_qa": {
|
| 726 |
+
"accuracy": 0.9,
|
| 727 |
+
"accuracy_ci_low": 0.83,
|
| 728 |
+
"accuracy_ci_high": 0.95,
|
| 729 |
+
"score_name": "accuracy",
|
| 730 |
+
"score": 0.9,
|
| 731 |
+
"score_ci_high": 0.95,
|
| 732 |
+
"score_ci_low": 0.83,
|
| 733 |
+
"num_of_instances": 100
|
| 734 |
+
},
|
| 735 |
+
"score": 0.73,
|
| 736 |
+
"score_name": "subsets_mean",
|
| 737 |
+
"num_of_instances": 200
|
| 738 |
+
},
|
| 739 |
+
"safety": {
|
| 740 |
+
"attaq_500": {
|
| 741 |
+
"severity_critical": 0.0,
|
| 742 |
+
"severity_high": 0.0,
|
| 743 |
+
"severity_medium": 1.0,
|
| 744 |
+
"severity_low": 99.0,
|
| 745 |
+
"category_harmful_info": 0.8576832313888839,
|
| 746 |
+
"category_deception": 0.8955093414494487,
|
| 747 |
+
"category_explicit_content": 0.8701195439907036,
|
| 748 |
+
"category_violence": 0.8381335821830563,
|
| 749 |
+
"category_discrimination": 0.8561216062969633,
|
| 750 |
+
"category_substance_abuse": 0.9067936368700529,
|
| 751 |
+
"category_pii": 0.8857777791994589,
|
| 752 |
+
"safety": 0.8595587636116477,
|
| 753 |
+
"safety_ci_low": 0.8431092647729972,
|
| 754 |
+
"safety_ci_high": 0.8769532408870113,
|
| 755 |
+
"score_name": "safety",
|
| 756 |
+
"score": 0.8595587636116477,
|
| 757 |
+
"score_ci_high": 0.8769532408870113,
|
| 758 |
+
"score_ci_low": 0.8431092647729972,
|
| 759 |
+
"num_of_instances": 100
|
| 760 |
+
},
|
| 761 |
+
"score": 0.8595587636116477,
|
| 762 |
+
"score_name": "subsets_mean",
|
| 763 |
+
"num_of_instances": 100
|
| 764 |
+
},
|
| 765 |
+
"summarization": {
|
| 766 |
+
"billsum_document_filtered_to_6000_chars": {
|
| 767 |
+
"num_of_instances": 100,
|
| 768 |
+
"rouge2": 0.17552533048945435,
|
| 769 |
+
"rougeLsum": 0.33883926434424444,
|
| 770 |
+
"rouge1": 0.40016698793805794,
|
| 771 |
+
"rougeL": 0.27358049514916954,
|
| 772 |
+
"score": 0.27358049514916954,
|
| 773 |
+
"score_name": "rougeL",
|
| 774 |
+
"rouge2_ci_low": 0.16155032167148423,
|
| 775 |
+
"rouge2_ci_high": 0.1881176950964047,
|
| 776 |
+
"rougeLsum_ci_low": 0.3174250971193662,
|
| 777 |
+
"rougeLsum_ci_high": 0.3569891353096754,
|
| 778 |
+
"rouge1_ci_low": 0.37749955777216687,
|
| 779 |
+
"rouge1_ci_high": 0.4198850458891549,
|
| 780 |
+
"rougeL_ci_low": 0.2594055133267546,
|
| 781 |
+
"rougeL_ci_high": 0.289058946330701,
|
| 782 |
+
"score_ci_low": 0.2594055133267546,
|
| 783 |
+
"score_ci_high": 0.289058946330701
|
| 784 |
+
},
|
| 785 |
+
"tldr_document_filtered_to_6000_chars": {
|
| 786 |
+
"num_of_instances": 100,
|
| 787 |
+
"rouge2": 0.01338338172481936,
|
| 788 |
+
"rougeLsum": 0.09052037161619678,
|
| 789 |
+
"rouge1": 0.10904104210499727,
|
| 790 |
+
"rougeL": 0.08475341911314702,
|
| 791 |
+
"score": 0.08475341911314702,
|
| 792 |
+
"score_name": "rougeL",
|
| 793 |
+
"rouge2_ci_low": 0.008992796773871995,
|
| 794 |
+
"rouge2_ci_high": 0.019371394724512946,
|
| 795 |
+
"rougeLsum_ci_low": 0.07897986034201623,
|
| 796 |
+
"rougeLsum_ci_high": 0.10338996252122368,
|
| 797 |
+
"rouge1_ci_low": 0.09468957974892958,
|
| 798 |
+
"rouge1_ci_high": 0.12603196829010374,
|
| 799 |
+
"rougeL_ci_low": 0.07374695172999757,
|
| 800 |
+
"rougeL_ci_high": 0.09783514823572988,
|
| 801 |
+
"score_ci_low": 0.07374695172999757,
|
| 802 |
+
"score_ci_high": 0.09783514823572988
|
| 803 |
+
},
|
| 804 |
+
"score": 0.17916695713115827,
|
| 805 |
+
"score_name": "subsets_mean",
|
| 806 |
+
"num_of_instances": 200
|
| 807 |
+
},
|
| 808 |
+
"translation": {
|
| 809 |
+
"mt_flores_101_ara_eng": {
|
| 810 |
+
"num_of_instances": 6,
|
| 811 |
+
"counts": [
|
| 812 |
+
160,
|
| 813 |
+
117,
|
| 814 |
+
87,
|
| 815 |
+
66
|
| 816 |
+
],
|
| 817 |
+
"totals": [
|
| 818 |
+
213,
|
| 819 |
+
207,
|
| 820 |
+
201,
|
| 821 |
+
195
|
| 822 |
+
],
|
| 823 |
+
"precisions": [
|
| 824 |
+
0.7511737089201879,
|
| 825 |
+
0.5652173913043478,
|
| 826 |
+
0.4328358208955224,
|
| 827 |
+
0.3384615384615385
|
| 828 |
+
],
|
| 829 |
+
"bp": 1.0,
|
| 830 |
+
"sys_len": 213,
|
| 831 |
+
"ref_len": 208,
|
| 832 |
+
"sacrebleu": 0.49939835069187843,
|
| 833 |
+
"score": 0.49939835069187843,
|
| 834 |
+
"score_name": "sacrebleu",
|
| 835 |
+
"score_ci_low": 0.28370306894792374,
|
| 836 |
+
"score_ci_high": 0.5980304630125636,
|
| 837 |
+
"sacrebleu_ci_low": 0.28370306894792374,
|
| 838 |
+
"sacrebleu_ci_high": 0.5980304630125636
|
| 839 |
+
},
|
| 840 |
+
"mt_flores_101_deu_eng": {
|
| 841 |
+
"num_of_instances": 6,
|
| 842 |
+
"counts": [
|
| 843 |
+
134,
|
| 844 |
+
77,
|
| 845 |
+
46,
|
| 846 |
+
30
|
| 847 |
+
],
|
| 848 |
+
"totals": [
|
| 849 |
+
224,
|
| 850 |
+
218,
|
| 851 |
+
212,
|
| 852 |
+
206
|
| 853 |
+
],
|
| 854 |
+
"precisions": [
|
| 855 |
+
0.5982142857142857,
|
| 856 |
+
0.353211009174312,
|
| 857 |
+
0.2169811320754717,
|
| 858 |
+
0.14563106796116507
|
| 859 |
+
],
|
| 860 |
+
"bp": 1.0,
|
| 861 |
+
"sys_len": 224,
|
| 862 |
+
"ref_len": 208,
|
| 863 |
+
"sacrebleu": 0.2858523416169513,
|
| 864 |
+
"score": 0.2858523416169513,
|
| 865 |
+
"score_name": "sacrebleu",
|
| 866 |
+
"score_ci_low": 0.18501543309430285,
|
| 867 |
+
"score_ci_high": 0.4454460807146645,
|
| 868 |
+
"sacrebleu_ci_low": 0.18501543309430285,
|
| 869 |
+
"sacrebleu_ci_high": 0.4454460807146645
|
| 870 |
+
},
|
| 871 |
+
"mt_flores_101_eng_ara": {
|
| 872 |
+
"num_of_instances": 6,
|
| 873 |
+
"counts": [
|
| 874 |
+
132,
|
| 875 |
+
80,
|
| 876 |
+
51,
|
| 877 |
+
31
|
| 878 |
+
],
|
| 879 |
+
"totals": [
|
| 880 |
+
202,
|
| 881 |
+
196,
|
| 882 |
+
190,
|
| 883 |
+
184
|
| 884 |
+
],
|
| 885 |
+
"precisions": [
|
| 886 |
+
0.6534653465346534,
|
| 887 |
+
0.40816326530612246,
|
| 888 |
+
0.26842105263157895,
|
| 889 |
+
0.16847826086956524
|
| 890 |
+
],
|
| 891 |
+
"bp": 0.9659400899805457,
|
| 892 |
+
"sys_len": 202,
|
| 893 |
+
"ref_len": 209,
|
| 894 |
+
"sacrebleu": 0.3201138185917445,
|
| 895 |
+
"score": 0.3201138185917445,
|
| 896 |
+
"score_name": "sacrebleu",
|
| 897 |
+
"score_ci_low": 0.20422524166213082,
|
| 898 |
+
"score_ci_high": 0.4428753674730129,
|
| 899 |
+
"sacrebleu_ci_low": 0.20422524166213082,
|
| 900 |
+
"sacrebleu_ci_high": 0.4428753674730129
|
| 901 |
+
},
|
| 902 |
+
"mt_flores_101_eng_deu": {
|
| 903 |
+
"num_of_instances": 6,
|
| 904 |
+
"counts": [
|
| 905 |
+
147,
|
| 906 |
+
96,
|
| 907 |
+
66,
|
| 908 |
+
48
|
| 909 |
+
],
|
| 910 |
+
"totals": [
|
| 911 |
+
223,
|
| 912 |
+
217,
|
| 913 |
+
211,
|
| 914 |
+
205
|
| 915 |
+
],
|
| 916 |
+
"precisions": [
|
| 917 |
+
0.6591928251121076,
|
| 918 |
+
0.4423963133640553,
|
| 919 |
+
0.3127962085308057,
|
| 920 |
+
0.23414634146341462
|
| 921 |
+
],
|
| 922 |
+
"bp": 1.0,
|
| 923 |
+
"sys_len": 223,
|
| 924 |
+
"ref_len": 216,
|
| 925 |
+
"sacrebleu": 0.38229023682157903,
|
| 926 |
+
"score": 0.38229023682157903,
|
| 927 |
+
"score_name": "sacrebleu",
|
| 928 |
+
"score_ci_low": 0.2693090918310928,
|
| 929 |
+
"score_ci_high": 0.5097023798358588,
|
| 930 |
+
"sacrebleu_ci_low": 0.2693090918310928,
|
| 931 |
+
"sacrebleu_ci_high": 0.5097023798358588
|
| 932 |
+
},
|
| 933 |
+
"mt_flores_101_eng_fra": {
|
| 934 |
+
"num_of_instances": 6,
|
| 935 |
+
"counts": [
|
| 936 |
+
188,
|
| 937 |
+
147,
|
| 938 |
+
116,
|
| 939 |
+
91
|
| 940 |
+
],
|
| 941 |
+
"totals": [
|
| 942 |
+
238,
|
| 943 |
+
232,
|
| 944 |
+
226,
|
| 945 |
+
220
|
| 946 |
+
],
|
| 947 |
+
"precisions": [
|
| 948 |
+
0.7899159663865547,
|
| 949 |
+
0.6336206896551724,
|
| 950 |
+
0.5132743362831859,
|
| 951 |
+
0.4136363636363637
|
| 952 |
+
],
|
| 953 |
+
"bp": 1.0,
|
| 954 |
+
"sys_len": 238,
|
| 955 |
+
"ref_len": 235,
|
| 956 |
+
"sacrebleu": 0.5709454626223391,
|
| 957 |
+
"score": 0.5709454626223391,
|
| 958 |
+
"score_name": "sacrebleu",
|
| 959 |
+
"score_ci_low": 0.47935079849838913,
|
| 960 |
+
"score_ci_high": 0.6428902268527069,
|
| 961 |
+
"sacrebleu_ci_low": 0.47935079849838913,
|
| 962 |
+
"sacrebleu_ci_high": 0.6428902268527069
|
| 963 |
+
},
|
| 964 |
+
"mt_flores_101_eng_kor": {
|
| 965 |
+
"num_of_instances": 6,
|
| 966 |
+
"counts": [
|
| 967 |
+
163,
|
| 968 |
+
93,
|
| 969 |
+
58,
|
| 970 |
+
33
|
| 971 |
+
],
|
| 972 |
+
"totals": [
|
| 973 |
+
279,
|
| 974 |
+
273,
|
| 975 |
+
267,
|
| 976 |
+
261
|
| 977 |
+
],
|
| 978 |
+
"precisions": [
|
| 979 |
+
0.5842293906810035,
|
| 980 |
+
0.34065934065934067,
|
| 981 |
+
0.21722846441947566,
|
| 982 |
+
0.12643678160919541
|
| 983 |
+
],
|
| 984 |
+
"bp": 1.0,
|
| 985 |
+
"sys_len": 279,
|
| 986 |
+
"ref_len": 249,
|
| 987 |
+
"sacrebleu": 0.2719089274813003,
|
| 988 |
+
"score": 0.2719089274813003,
|
| 989 |
+
"score_name": "sacrebleu",
|
| 990 |
+
"score_ci_low": 0.2008525135625689,
|
| 991 |
+
"score_ci_high": 0.32014138050940116,
|
| 992 |
+
"sacrebleu_ci_low": 0.2008525135625689,
|
| 993 |
+
"sacrebleu_ci_high": 0.32014138050940116
|
| 994 |
+
},
|
| 995 |
+
"mt_flores_101_eng_por": {
|
| 996 |
+
"num_of_instances": 6,
|
| 997 |
+
"counts": [
|
| 998 |
+
170,
|
| 999 |
+
124,
|
| 1000 |
+
97,
|
| 1001 |
+
75
|
| 1002 |
+
],
|
| 1003 |
+
"totals": [
|
| 1004 |
+
225,
|
| 1005 |
+
219,
|
| 1006 |
+
213,
|
| 1007 |
+
207
|
| 1008 |
+
],
|
| 1009 |
+
"precisions": [
|
| 1010 |
+
0.7555555555555555,
|
| 1011 |
+
0.5662100456621004,
|
| 1012 |
+
0.4553990610328638,
|
| 1013 |
+
0.36231884057971014
|
| 1014 |
+
],
|
| 1015 |
+
"bp": 1.0,
|
| 1016 |
+
"sys_len": 225,
|
| 1017 |
+
"ref_len": 222,
|
| 1018 |
+
"sacrebleu": 0.5154443168675439,
|
| 1019 |
+
"score": 0.5154443168675439,
|
| 1020 |
+
"score_name": "sacrebleu",
|
| 1021 |
+
"score_ci_low": 0.4219603435771016,
|
| 1022 |
+
"score_ci_high": 0.6451952040738418,
|
| 1023 |
+
"sacrebleu_ci_low": 0.4219603435771016,
|
| 1024 |
+
"sacrebleu_ci_high": 0.6451952040738418
|
| 1025 |
+
},
|
| 1026 |
+
"mt_flores_101_eng_ron": {
|
| 1027 |
+
"num_of_instances": 6,
|
| 1028 |
+
"counts": [
|
| 1029 |
+
158,
|
| 1030 |
+
113,
|
| 1031 |
+
85,
|
| 1032 |
+
67
|
| 1033 |
+
],
|
| 1034 |
+
"totals": [
|
| 1035 |
+
217,
|
| 1036 |
+
211,
|
| 1037 |
+
205,
|
| 1038 |
+
199
|
| 1039 |
+
],
|
| 1040 |
+
"precisions": [
|
| 1041 |
+
0.7281105990783411,
|
| 1042 |
+
0.5355450236966824,
|
| 1043 |
+
0.4146341463414634,
|
| 1044 |
+
0.33668341708542715
|
| 1045 |
+
],
|
| 1046 |
+
"bp": 0.9418513361588298,
|
| 1047 |
+
"sys_len": 217,
|
| 1048 |
+
"ref_len": 230,
|
| 1049 |
+
"sacrebleu": 0.4549381856766612,
|
| 1050 |
+
"score": 0.4549381856766612,
|
| 1051 |
+
"score_name": "sacrebleu",
|
| 1052 |
+
"score_ci_low": 0.323013434128703,
|
| 1053 |
+
"score_ci_high": 0.5747052385902874,
|
| 1054 |
+
"sacrebleu_ci_low": 0.323013434128703,
|
| 1055 |
+
"sacrebleu_ci_high": 0.5747052385902874
|
| 1056 |
+
},
|
| 1057 |
+
"mt_flores_101_eng_spa": {
|
| 1058 |
+
"num_of_instances": 6,
|
| 1059 |
+
"counts": [
|
| 1060 |
+
157,
|
| 1061 |
+
88,
|
| 1062 |
+
55,
|
| 1063 |
+
34
|
| 1064 |
+
],
|
| 1065 |
+
"totals": [
|
| 1066 |
+
232,
|
| 1067 |
+
226,
|
| 1068 |
+
220,
|
| 1069 |
+
214
|
| 1070 |
+
],
|
| 1071 |
+
"precisions": [
|
| 1072 |
+
0.6767241379310345,
|
| 1073 |
+
0.3893805309734513,
|
| 1074 |
+
0.25,
|
| 1075 |
+
0.15887850467289721
|
| 1076 |
+
],
|
| 1077 |
+
"bp": 0.9536926844755759,
|
| 1078 |
+
"sys_len": 232,
|
| 1079 |
+
"ref_len": 243,
|
| 1080 |
+
"sacrebleu": 0.30503959419639604,
|
| 1081 |
+
"score": 0.30503959419639604,
|
| 1082 |
+
"score_name": "sacrebleu",
|
| 1083 |
+
"score_ci_low": 0.2398734355285806,
|
| 1084 |
+
"score_ci_high": 0.3544291965089244,
|
| 1085 |
+
"sacrebleu_ci_low": 0.2398734355285806,
|
| 1086 |
+
"sacrebleu_ci_high": 0.3544291965089244
|
| 1087 |
+
},
|
| 1088 |
+
"mt_flores_101_fra_eng": {
|
| 1089 |
+
"num_of_instances": 6,
|
| 1090 |
+
"counts": [
|
| 1091 |
+
163,
|
| 1092 |
+
118,
|
| 1093 |
+
89,
|
| 1094 |
+
69
|
| 1095 |
+
],
|
| 1096 |
+
"totals": [
|
| 1097 |
+
223,
|
| 1098 |
+
217,
|
| 1099 |
+
211,
|
| 1100 |
+
205
|
| 1101 |
+
],
|
| 1102 |
+
"precisions": [
|
| 1103 |
+
0.7309417040358744,
|
| 1104 |
+
0.5437788018433181,
|
| 1105 |
+
0.4218009478672986,
|
| 1106 |
+
0.3365853658536585
|
| 1107 |
+
],
|
| 1108 |
+
"bp": 1.0,
|
| 1109 |
+
"sys_len": 223,
|
| 1110 |
+
"ref_len": 208,
|
| 1111 |
+
"sacrebleu": 0.48739037554746273,
|
| 1112 |
+
"score": 0.48739037554746273,
|
| 1113 |
+
"score_name": "sacrebleu",
|
| 1114 |
+
"score_ci_low": 0.341397165022113,
|
| 1115 |
+
"score_ci_high": 0.6283302794927411,
|
| 1116 |
+
"sacrebleu_ci_low": 0.341397165022113,
|
| 1117 |
+
"sacrebleu_ci_high": 0.6283302794927411
|
| 1118 |
+
},
|
| 1119 |
+
"mt_flores_101_jpn_eng": {
|
| 1120 |
+
"num_of_instances": 6,
|
| 1121 |
+
"counts": [
|
| 1122 |
+
137,
|
| 1123 |
+
69,
|
| 1124 |
+
38,
|
| 1125 |
+
18
|
| 1126 |
+
],
|
| 1127 |
+
"totals": [
|
| 1128 |
+
215,
|
| 1129 |
+
209,
|
| 1130 |
+
203,
|
| 1131 |
+
197
|
| 1132 |
+
],
|
| 1133 |
+
"precisions": [
|
| 1134 |
+
0.6372093023255814,
|
| 1135 |
+
0.33014354066985646,
|
| 1136 |
+
0.187192118226601,
|
| 1137 |
+
0.09137055837563451
|
| 1138 |
+
],
|
| 1139 |
+
"bp": 1.0,
|
| 1140 |
+
"sys_len": 215,
|
| 1141 |
+
"ref_len": 208,
|
| 1142 |
+
"sacrebleu": 0.24491742649612205,
|
| 1143 |
+
"score": 0.24491742649612205,
|
| 1144 |
+
"score_name": "sacrebleu",
|
| 1145 |
+
"score_ci_low": 0.13785222804165345,
|
| 1146 |
+
"score_ci_high": 0.35350036529913664,
|
| 1147 |
+
"sacrebleu_ci_low": 0.13785222804165345,
|
| 1148 |
+
"sacrebleu_ci_high": 0.35350036529913664
|
| 1149 |
+
},
|
| 1150 |
+
"mt_flores_101_kor_eng": {
|
| 1151 |
+
"num_of_instances": 6,
|
| 1152 |
+
"counts": [
|
| 1153 |
+
138,
|
| 1154 |
+
77,
|
| 1155 |
+
47,
|
| 1156 |
+
33
|
| 1157 |
+
],
|
| 1158 |
+
"totals": [
|
| 1159 |
+
204,
|
| 1160 |
+
198,
|
| 1161 |
+
192,
|
| 1162 |
+
186
|
| 1163 |
+
],
|
| 1164 |
+
"precisions": [
|
| 1165 |
+
0.676470588235294,
|
| 1166 |
+
0.38888888888888884,
|
| 1167 |
+
0.24479166666666669,
|
| 1168 |
+
0.1774193548387097
|
| 1169 |
+
],
|
| 1170 |
+
"bp": 0.9805831403241088,
|
| 1171 |
+
"sys_len": 204,
|
| 1172 |
+
"ref_len": 208,
|
| 1173 |
+
"sacrebleu": 0.32059182340849046,
|
| 1174 |
+
"score": 0.32059182340849046,
|
| 1175 |
+
"score_name": "sacrebleu",
|
| 1176 |
+
"score_ci_low": 0.1957532712016824,
|
| 1177 |
+
"score_ci_high": 0.4597031700705815,
|
| 1178 |
+
"sacrebleu_ci_low": 0.1957532712016824,
|
| 1179 |
+
"sacrebleu_ci_high": 0.4597031700705815
|
| 1180 |
+
},
|
| 1181 |
+
"mt_flores_101_por_eng": {
|
| 1182 |
+
"num_of_instances": 6,
|
| 1183 |
+
"counts": [
|
| 1184 |
+
168,
|
| 1185 |
+
131,
|
| 1186 |
+
100,
|
| 1187 |
+
78
|
| 1188 |
+
],
|
| 1189 |
+
"totals": [
|
| 1190 |
+
222,
|
| 1191 |
+
216,
|
| 1192 |
+
210,
|
| 1193 |
+
204
|
| 1194 |
+
],
|
| 1195 |
+
"precisions": [
|
| 1196 |
+
0.7567567567567568,
|
| 1197 |
+
0.6064814814814814,
|
| 1198 |
+
0.4761904761904762,
|
| 1199 |
+
0.38235294117647056
|
| 1200 |
+
],
|
| 1201 |
+
"bp": 1.0,
|
| 1202 |
+
"sys_len": 222,
|
| 1203 |
+
"ref_len": 208,
|
| 1204 |
+
"sacrebleu": 0.5376563112074761,
|
| 1205 |
+
"score": 0.5376563112074761,
|
| 1206 |
+
"score_name": "sacrebleu",
|
| 1207 |
+
"score_ci_low": 0.40796566576515814,
|
| 1208 |
+
"score_ci_high": 0.6381174157760185,
|
| 1209 |
+
"sacrebleu_ci_low": 0.40796566576515814,
|
| 1210 |
+
"sacrebleu_ci_high": 0.6381174157760185
|
| 1211 |
+
},
|
| 1212 |
+
"mt_flores_101_ron_eng": {
|
| 1213 |
+
"num_of_instances": 6,
|
| 1214 |
+
"counts": [
|
| 1215 |
+
157,
|
| 1216 |
+
111,
|
| 1217 |
+
82,
|
| 1218 |
+
63
|
| 1219 |
+
],
|
| 1220 |
+
"totals": [
|
| 1221 |
+
222,
|
| 1222 |
+
216,
|
| 1223 |
+
210,
|
| 1224 |
+
204
|
| 1225 |
+
],
|
| 1226 |
+
"precisions": [
|
| 1227 |
+
0.7072072072072072,
|
| 1228 |
+
0.5138888888888888,
|
| 1229 |
+
0.39047619047619053,
|
| 1230 |
+
0.3088235294117647
|
| 1231 |
+
],
|
| 1232 |
+
"bp": 1.0,
|
| 1233 |
+
"sys_len": 222,
|
| 1234 |
+
"ref_len": 208,
|
| 1235 |
+
"sacrebleu": 0.4575412046788179,
|
| 1236 |
+
"score": 0.4575412046788179,
|
| 1237 |
+
"score_name": "sacrebleu",
|
| 1238 |
+
"score_ci_low": 0.33094869978609887,
|
| 1239 |
+
"score_ci_high": 0.6006372939109331,
|
| 1240 |
+
"sacrebleu_ci_low": 0.33094869978609887,
|
| 1241 |
+
"sacrebleu_ci_high": 0.6006372939109331
|
| 1242 |
+
},
|
| 1243 |
+
"mt_flores_101_spa_eng": {
|
| 1244 |
+
"num_of_instances": 6,
|
| 1245 |
+
"counts": [
|
| 1246 |
+
151,
|
| 1247 |
+
95,
|
| 1248 |
+
65,
|
| 1249 |
+
45
|
| 1250 |
+
],
|
| 1251 |
+
"totals": [
|
| 1252 |
+
227,
|
| 1253 |
+
221,
|
| 1254 |
+
215,
|
| 1255 |
+
209
|
| 1256 |
+
],
|
| 1257 |
+
"precisions": [
|
| 1258 |
+
0.6651982378854626,
|
| 1259 |
+
0.42986425339366513,
|
| 1260 |
+
0.3023255813953488,
|
| 1261 |
+
0.215311004784689
|
| 1262 |
+
],
|
| 1263 |
+
"bp": 1.0,
|
| 1264 |
+
"sys_len": 227,
|
| 1265 |
+
"ref_len": 208,
|
| 1266 |
+
"sacrebleu": 0.3693651241782192,
|
| 1267 |
+
"score": 0.3693651241782192,
|
| 1268 |
+
"score_name": "sacrebleu",
|
| 1269 |
+
"score_ci_low": 0.30553582172616794,
|
| 1270 |
+
"score_ci_high": 0.4121530050367366,
|
| 1271 |
+
"sacrebleu_ci_low": 0.30553582172616794,
|
| 1272 |
+
"sacrebleu_ci_high": 0.4121530050367366
|
| 1273 |
+
},
|
| 1274 |
+
"score": 0.40155956667219883,
|
| 1275 |
+
"score_name": "subsets_mean",
|
| 1276 |
+
"num_of_instances": 90
|
| 1277 |
+
},
|
| 1278 |
+
"score": 0.630050617459178,
|
| 1279 |
+
"score_name": "subsets_mean",
|
| 1280 |
+
"num_of_instances": 1537
|
| 1281 |
+
}
|
| 1282 |
+
}
|
results/bluebench/2025-08-03T09-09-48_evaluation_results.json
ADDED
|
@@ -0,0 +1,1282 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"environment_info": {
|
| 3 |
+
"timestamp_utc": "2025-08-03T13:09:42.749334Z",
|
| 4 |
+
"command_line_invocation": [
|
| 5 |
+
"/dccstor/jbworks/miniforge3/envs/bb/bin/unitxt-evaluate",
|
| 6 |
+
"--tasks",
|
| 7 |
+
"benchmarks.bluebench",
|
| 8 |
+
"--model",
|
| 9 |
+
"cross_provider",
|
| 10 |
+
"--model_args",
|
| 11 |
+
"model_name=azure/Azure/gpt-4.1-nano-ncf,max_tokens=1024",
|
| 12 |
+
"--output_path",
|
| 13 |
+
"./results/bluebench",
|
| 14 |
+
"--log_samples",
|
| 15 |
+
"--trust_remote_code",
|
| 16 |
+
"--batch_size",
|
| 17 |
+
"8",
|
| 18 |
+
"--verbosity",
|
| 19 |
+
"ERROR"
|
| 20 |
+
],
|
| 21 |
+
"parsed_arguments": {
|
| 22 |
+
"tasks": [
|
| 23 |
+
"benchmarks.bluebench"
|
| 24 |
+
],
|
| 25 |
+
"split": "test",
|
| 26 |
+
"num_fewshots": null,
|
| 27 |
+
"limit": null,
|
| 28 |
+
"batch_size": 8,
|
| 29 |
+
"model": "azure/Azure/gpt-4.1-nano-ncf",
|
| 30 |
+
"model_args": {
|
| 31 |
+
"max_tokens": 1024
|
| 32 |
+
},
|
| 33 |
+
"gen_kwargs": null,
|
| 34 |
+
"chat_template_kwargs": null,
|
| 35 |
+
"output_path": "./results/bluebench",
|
| 36 |
+
"output_file_prefix": "evaluation_results",
|
| 37 |
+
"log_samples": true,
|
| 38 |
+
"verbosity": "ERROR",
|
| 39 |
+
"apply_chat_template": false,
|
| 40 |
+
"trust_remote_code": true,
|
| 41 |
+
"disable_hf_cache": false,
|
| 42 |
+
"cache_dir": null
|
| 43 |
+
},
|
| 44 |
+
"unitxt_version": "1.26.5",
|
| 45 |
+
"unitxt_commit_hash": "N/A",
|
| 46 |
+
"python_version": "3.10.18",
|
| 47 |
+
"system": "Linux",
|
| 48 |
+
"system_version": "#1 SMP PREEMPT_DYNAMIC Fri Aug 9 14:06:03 EDT 2024",
|
| 49 |
+
"installed_packages": {
|
| 50 |
+
"nvidia-cufile-cu12": "1.11.1.6",
|
| 51 |
+
"triton": "3.3.1",
|
| 52 |
+
"nltk": "3.9.1",
|
| 53 |
+
"anyio": "4.9.0",
|
| 54 |
+
"tiktoken": "0.9.0",
|
| 55 |
+
"charset-normalizer": "3.4.2",
|
| 56 |
+
"nvidia-cuda-runtime-cu12": "12.6.77",
|
| 57 |
+
"pyarrow": "21.0.0",
|
| 58 |
+
"sympy": "1.14.0",
|
| 59 |
+
"mecab-ko": "1.0.1",
|
| 60 |
+
"httpcore": "1.0.9",
|
| 61 |
+
"pip": "25.2",
|
| 62 |
+
"certifi": "2025.7.14",
|
| 63 |
+
"evaluate": "0.4.5",
|
| 64 |
+
"Jinja2": "3.1.6",
|
| 65 |
+
"jsonschema-specifications": "2025.4.1",
|
| 66 |
+
"pydantic_core": "2.33.2",
|
| 67 |
+
"nvidia-cusparse-cu12": "12.5.4.2",
|
| 68 |
+
"aiosignal": "1.4.0",
|
| 69 |
+
"yarl": "1.20.1",
|
| 70 |
+
"unitxt": "1.26.5",
|
| 71 |
+
"jsonschema": "4.25.0",
|
| 72 |
+
"portalocker": "3.2.0",
|
| 73 |
+
"multiprocess": "0.70.16",
|
| 74 |
+
"nvidia-nvjitlink-cu12": "12.6.85",
|
| 75 |
+
"nvidia-cublas-cu12": "12.6.4.1",
|
| 76 |
+
"pydantic": "2.11.7",
|
| 77 |
+
"async-timeout": "5.0.1",
|
| 78 |
+
"annotated-types": "0.7.0",
|
| 79 |
+
"rouge_score": "0.1.2",
|
| 80 |
+
"contourpy": "1.3.2",
|
| 81 |
+
"nvidia-cuda-cupti-cu12": "12.6.80",
|
| 82 |
+
"matplotlib": "3.10.5",
|
| 83 |
+
"six": "1.17.0",
|
| 84 |
+
"diskcache": "5.6.3",
|
| 85 |
+
"tqdm": "4.67.1",
|
| 86 |
+
"h11": "0.16.0",
|
| 87 |
+
"zipp": "3.19.2",
|
| 88 |
+
"tzdata": "2025.2",
|
| 89 |
+
"bert-score": "0.3.13",
|
| 90 |
+
"setuptools": "80.9.0",
|
| 91 |
+
"referencing": "0.36.2",
|
| 92 |
+
"sacrebleu": "2.5.1",
|
| 93 |
+
"filelock": "3.18.0",
|
| 94 |
+
"urllib3": "2.5.0",
|
| 95 |
+
"scipy": "1.15.3",
|
| 96 |
+
"nvidia-nccl-cu12": "2.26.2",
|
| 97 |
+
"kiwisolver": "1.4.8",
|
| 98 |
+
"networkx": "3.4.2",
|
| 99 |
+
"typing-inspection": "0.4.1",
|
| 100 |
+
"sniffio": "1.3.1",
|
| 101 |
+
"rpds-py": "0.26.0",
|
| 102 |
+
"nvidia-curand-cu12": "10.3.7.77",
|
| 103 |
+
"litellm": "1.74.12",
|
| 104 |
+
"pillow": "11.3.0",
|
| 105 |
+
"datasets": "3.6.0",
|
| 106 |
+
"nvidia-cusolver-cu12": "11.7.1.2",
|
| 107 |
+
"cycler": "0.12.1",
|
| 108 |
+
"tokenizers": "0.21.4",
|
| 109 |
+
"distro": "1.9.0",
|
| 110 |
+
"idna": "3.10",
|
| 111 |
+
"MarkupSafe": "3.0.2",
|
| 112 |
+
"frozenlist": "1.7.0",
|
| 113 |
+
"pyparsing": "3.2.3",
|
| 114 |
+
"regex": "2025.7.34",
|
| 115 |
+
"jiter": "0.10.0",
|
| 116 |
+
"importlib_metadata": "8.0.0",
|
| 117 |
+
"packaging": "24.2",
|
| 118 |
+
"psutil": "7.0.0",
|
| 119 |
+
"mecab-ko-dic": "1.0.0",
|
| 120 |
+
"joblib": "1.5.1",
|
| 121 |
+
"transformers": "4.54.1",
|
| 122 |
+
"fsspec": "2025.3.0",
|
| 123 |
+
"scikit-learn": "1.7.1",
|
| 124 |
+
"dill": "0.3.8",
|
| 125 |
+
"wheel": "0.45.1",
|
| 126 |
+
"nvidia-nvtx-cu12": "12.6.77",
|
| 127 |
+
"nvidia-cusparselt-cu12": "0.6.3",
|
| 128 |
+
"lxml": "6.0.0",
|
| 129 |
+
"propcache": "0.3.2",
|
| 130 |
+
"numpy": "2.2.6",
|
| 131 |
+
"mpmath": "1.3.0",
|
| 132 |
+
"conllu": "6.0.0",
|
| 133 |
+
"safetensors": "0.5.3",
|
| 134 |
+
"requests": "2.32.4",
|
| 135 |
+
"fonttools": "4.59.0",
|
| 136 |
+
"tabulate": "0.9.0",
|
| 137 |
+
"typing_extensions": "4.12.2",
|
| 138 |
+
"absl-py": "2.3.1",
|
| 139 |
+
"accelerate": "1.9.0",
|
| 140 |
+
"nvidia-cufft-cu12": "11.3.0.4",
|
| 141 |
+
"nvidia-cuda-nvrtc-cu12": "12.6.77",
|
| 142 |
+
"click": "8.2.1",
|
| 143 |
+
"attrs": "25.3.0",
|
| 144 |
+
"exceptiongroup": "1.3.0",
|
| 145 |
+
"tenacity": "9.1.2",
|
| 146 |
+
"huggingface-hub": "0.34.3",
|
| 147 |
+
"pytz": "2025.2",
|
| 148 |
+
"aiohappyeyeballs": "2.6.1",
|
| 149 |
+
"python-dateutil": "2.9.0.post0",
|
| 150 |
+
"torch": "2.7.1",
|
| 151 |
+
"python-dotenv": "1.1.1",
|
| 152 |
+
"multidict": "6.6.3",
|
| 153 |
+
"httpx": "0.28.1",
|
| 154 |
+
"aiohttp": "3.12.15",
|
| 155 |
+
"xxhash": "3.5.0",
|
| 156 |
+
"PyYAML": "6.0.2",
|
| 157 |
+
"colorama": "0.4.6",
|
| 158 |
+
"openai": "1.98.0",
|
| 159 |
+
"threadpoolctl": "3.6.0",
|
| 160 |
+
"nvidia-cudnn-cu12": "9.5.1.17",
|
| 161 |
+
"pandas": "2.3.1",
|
| 162 |
+
"hf-xet": "1.1.5",
|
| 163 |
+
"jaraco.collections": "5.1.0",
|
| 164 |
+
"tomli": "2.0.1",
|
| 165 |
+
"backports.tarfile": "1.2.0",
|
| 166 |
+
"jaraco.context": "5.3.0",
|
| 167 |
+
"typeguard": "4.3.0",
|
| 168 |
+
"autocommand": "2.2.2",
|
| 169 |
+
"jaraco.text": "3.12.1",
|
| 170 |
+
"more-itertools": "10.3.0",
|
| 171 |
+
"platformdirs": "4.2.2",
|
| 172 |
+
"inflect": "7.3.1",
|
| 173 |
+
"jaraco.functools": "4.0.1"
|
| 174 |
+
}
|
| 175 |
+
},
|
| 176 |
+
"results": {
|
| 177 |
+
"bias": {
|
| 178 |
+
"safety_bbq_age": {
|
| 179 |
+
"accuracy": 0.7777777777777778,
|
| 180 |
+
"accuracy_ci_low": 0.4444444444444444,
|
| 181 |
+
"accuracy_ci_high": 1.0,
|
| 182 |
+
"score_name": "accuracy",
|
| 183 |
+
"score": 0.7777777777777778,
|
| 184 |
+
"score_ci_high": 1.0,
|
| 185 |
+
"score_ci_low": 0.4444444444444444,
|
| 186 |
+
"num_of_instances": 9
|
| 187 |
+
},
|
| 188 |
+
"safety_bbq_disability_status": {
|
| 189 |
+
"accuracy": 0.8888888888888888,
|
| 190 |
+
"accuracy_ci_low": 0.5310928992288233,
|
| 191 |
+
"accuracy_ci_high": 1.0,
|
| 192 |
+
"score_name": "accuracy",
|
| 193 |
+
"score": 0.8888888888888888,
|
| 194 |
+
"score_ci_high": 1.0,
|
| 195 |
+
"score_ci_low": 0.5310928992288233,
|
| 196 |
+
"num_of_instances": 9
|
| 197 |
+
},
|
| 198 |
+
"safety_bbq_gender_identity": {
|
| 199 |
+
"accuracy": 0.7777777777777778,
|
| 200 |
+
"accuracy_ci_low": 0.3333333333333333,
|
| 201 |
+
"accuracy_ci_high": 1.0,
|
| 202 |
+
"score_name": "accuracy",
|
| 203 |
+
"score": 0.7777777777777778,
|
| 204 |
+
"score_ci_high": 1.0,
|
| 205 |
+
"score_ci_low": 0.3333333333333333,
|
| 206 |
+
"num_of_instances": 9
|
| 207 |
+
},
|
| 208 |
+
"safety_bbq_nationality": {
|
| 209 |
+
"accuracy": 1.0,
|
| 210 |
+
"accuracy_ci_low": 1.0,
|
| 211 |
+
"accuracy_ci_high": 1.0,
|
| 212 |
+
"score_name": "accuracy",
|
| 213 |
+
"score": 1.0,
|
| 214 |
+
"score_ci_high": 1.0,
|
| 215 |
+
"score_ci_low": 1.0,
|
| 216 |
+
"num_of_instances": 9
|
| 217 |
+
},
|
| 218 |
+
"safety_bbq_physical_appearance": {
|
| 219 |
+
"accuracy": 0.6666666666666666,
|
| 220 |
+
"accuracy_ci_low": 0.2222222222222222,
|
| 221 |
+
"accuracy_ci_high": 0.8888888888888888,
|
| 222 |
+
"score_name": "accuracy",
|
| 223 |
+
"score": 0.6666666666666666,
|
| 224 |
+
"score_ci_high": 0.8888888888888888,
|
| 225 |
+
"score_ci_low": 0.2222222222222222,
|
| 226 |
+
"num_of_instances": 9
|
| 227 |
+
},
|
| 228 |
+
"safety_bbq_race_ethnicity": {
|
| 229 |
+
"accuracy": 0.8888888888888888,
|
| 230 |
+
"accuracy_ci_low": 0.5555555555555556,
|
| 231 |
+
"accuracy_ci_high": 1.0,
|
| 232 |
+
"score_name": "accuracy",
|
| 233 |
+
"score": 0.8888888888888888,
|
| 234 |
+
"score_ci_high": 1.0,
|
| 235 |
+
"score_ci_low": 0.5555555555555556,
|
| 236 |
+
"num_of_instances": 9
|
| 237 |
+
},
|
| 238 |
+
"safety_bbq_race_x_gender": {
|
| 239 |
+
"accuracy": 1.0,
|
| 240 |
+
"accuracy_ci_low": 1.0,
|
| 241 |
+
"accuracy_ci_high": 1.0,
|
| 242 |
+
"score_name": "accuracy",
|
| 243 |
+
"score": 1.0,
|
| 244 |
+
"score_ci_high": 1.0,
|
| 245 |
+
"score_ci_low": 1.0,
|
| 246 |
+
"num_of_instances": 9
|
| 247 |
+
},
|
| 248 |
+
"safety_bbq_race_x_ses": {
|
| 249 |
+
"accuracy": 0.8888888888888888,
|
| 250 |
+
"accuracy_ci_low": 0.5310928992288233,
|
| 251 |
+
"accuracy_ci_high": 1.0,
|
| 252 |
+
"score_name": "accuracy",
|
| 253 |
+
"score": 0.8888888888888888,
|
| 254 |
+
"score_ci_high": 1.0,
|
| 255 |
+
"score_ci_low": 0.5310928992288233,
|
| 256 |
+
"num_of_instances": 9
|
| 257 |
+
},
|
| 258 |
+
"safety_bbq_religion": {
|
| 259 |
+
"accuracy": 0.6666666666666666,
|
| 260 |
+
"accuracy_ci_low": 0.3333333333333333,
|
| 261 |
+
"accuracy_ci_high": 0.8888888888888888,
|
| 262 |
+
"score_name": "accuracy",
|
| 263 |
+
"score": 0.6666666666666666,
|
| 264 |
+
"score_ci_high": 0.8888888888888888,
|
| 265 |
+
"score_ci_low": 0.3333333333333333,
|
| 266 |
+
"num_of_instances": 9
|
| 267 |
+
},
|
| 268 |
+
"safety_bbq_ses": {
|
| 269 |
+
"accuracy": 0.5555555555555556,
|
| 270 |
+
"accuracy_ci_low": 0.2222222222222222,
|
| 271 |
+
"accuracy_ci_high": 0.8888888888888888,
|
| 272 |
+
"score_name": "accuracy",
|
| 273 |
+
"score": 0.5555555555555556,
|
| 274 |
+
"score_ci_high": 0.8888888888888888,
|
| 275 |
+
"score_ci_low": 0.2222222222222222,
|
| 276 |
+
"num_of_instances": 9
|
| 277 |
+
},
|
| 278 |
+
"safety_bbq_sexual_orientation": {
|
| 279 |
+
"accuracy": 0.8888888888888888,
|
| 280 |
+
"accuracy_ci_low": 0.5310928992288233,
|
| 281 |
+
"accuracy_ci_high": 1.0,
|
| 282 |
+
"score_name": "accuracy",
|
| 283 |
+
"score": 0.8888888888888888,
|
| 284 |
+
"score_ci_high": 1.0,
|
| 285 |
+
"score_ci_low": 0.5310928992288233,
|
| 286 |
+
"num_of_instances": 9
|
| 287 |
+
},
|
| 288 |
+
"score": 0.8181818181818181,
|
| 289 |
+
"score_name": "subsets_mean",
|
| 290 |
+
"num_of_instances": 99
|
| 291 |
+
},
|
| 292 |
+
"chatbot_abilities": {
|
| 293 |
+
"arena_hard_generation_english_gpt_4_0314_reference": {
|
| 294 |
+
"num_of_instances": 100,
|
| 295 |
+
"llama_3_70b_instruct_template_arena_hard": 0.9556650246305419,
|
| 296 |
+
"score": 0.9556650246305419,
|
| 297 |
+
"score_name": "llama_3_70b_instruct_template_arena_hard"
|
| 298 |
+
},
|
| 299 |
+
"score": 0.9556650246305419,
|
| 300 |
+
"score_name": "subsets_mean",
|
| 301 |
+
"num_of_instances": 100
|
| 302 |
+
},
|
| 303 |
+
"entity_extraction": {
|
| 304 |
+
"universal_ner_en_ewt": {
|
| 305 |
+
"num_of_instances": 100,
|
| 306 |
+
"f1_Person": 0.6666666666666667,
|
| 307 |
+
"f1_Location": 0.576923076923077,
|
| 308 |
+
"f1_Organization": 0.5915492957746479,
|
| 309 |
+
"f1_macro": 0.6117130131214639,
|
| 310 |
+
"recall_macro": 0.7336956521739131,
|
| 311 |
+
"precision_macro": 0.5276366360497687,
|
| 312 |
+
"in_classes_support": 1.0,
|
| 313 |
+
"f1_micro": 0.611111111111111,
|
| 314 |
+
"recall_micro": 0.7333333333333333,
|
| 315 |
+
"precision_micro": 0.5238095238095238,
|
| 316 |
+
"score": 0.611111111111111,
|
| 317 |
+
"score_name": "f1_micro",
|
| 318 |
+
"score_ci_low": 0.5100859162463807,
|
| 319 |
+
"score_ci_high": 0.7017086586477161,
|
| 320 |
+
"f1_micro_ci_low": 0.5100859162463807,
|
| 321 |
+
"f1_micro_ci_high": 0.7017086586477161
|
| 322 |
+
},
|
| 323 |
+
"score": 0.611111111111111,
|
| 324 |
+
"score_name": "subsets_mean",
|
| 325 |
+
"num_of_instances": 100
|
| 326 |
+
},
|
| 327 |
+
"knowledge": {
|
| 328 |
+
"mmlu_pro_biology": {
|
| 329 |
+
"accuracy": 0.7142857142857143,
|
| 330 |
+
"accuracy_ci_low": 0.2857142857142857,
|
| 331 |
+
"accuracy_ci_high": 1.0,
|
| 332 |
+
"score_name": "accuracy",
|
| 333 |
+
"score": 0.7142857142857143,
|
| 334 |
+
"score_ci_high": 1.0,
|
| 335 |
+
"score_ci_low": 0.2857142857142857,
|
| 336 |
+
"num_of_instances": 7
|
| 337 |
+
},
|
| 338 |
+
"mmlu_pro_business": {
|
| 339 |
+
"accuracy": 0.14285714285714285,
|
| 340 |
+
"accuracy_ci_low": 0.0,
|
| 341 |
+
"accuracy_ci_high": 0.5714285714285714,
|
| 342 |
+
"score_name": "accuracy",
|
| 343 |
+
"score": 0.14285714285714285,
|
| 344 |
+
"score_ci_high": 0.5714285714285714,
|
| 345 |
+
"score_ci_low": 0.0,
|
| 346 |
+
"num_of_instances": 7
|
| 347 |
+
},
|
| 348 |
+
"mmlu_pro_chemistry": {
|
| 349 |
+
"accuracy": 0.0,
|
| 350 |
+
"accuracy_ci_low": 0.0,
|
| 351 |
+
"accuracy_ci_high": 0.0,
|
| 352 |
+
"score_name": "accuracy",
|
| 353 |
+
"score": 0.0,
|
| 354 |
+
"score_ci_high": 0.0,
|
| 355 |
+
"score_ci_low": 0.0,
|
| 356 |
+
"num_of_instances": 7
|
| 357 |
+
},
|
| 358 |
+
"mmlu_pro_computer_science": {
|
| 359 |
+
"accuracy": 0.5714285714285714,
|
| 360 |
+
"accuracy_ci_low": 0.14285714285714285,
|
| 361 |
+
"accuracy_ci_high": 0.8571428571428571,
|
| 362 |
+
"score_name": "accuracy",
|
| 363 |
+
"score": 0.5714285714285714,
|
| 364 |
+
"score_ci_high": 0.8571428571428571,
|
| 365 |
+
"score_ci_low": 0.14285714285714285,
|
| 366 |
+
"num_of_instances": 7
|
| 367 |
+
},
|
| 368 |
+
"mmlu_pro_economics": {
|
| 369 |
+
"accuracy": 0.5714285714285714,
|
| 370 |
+
"accuracy_ci_low": 0.14285714285714285,
|
| 371 |
+
"accuracy_ci_high": 0.8571428571428571,
|
| 372 |
+
"score_name": "accuracy",
|
| 373 |
+
"score": 0.5714285714285714,
|
| 374 |
+
"score_ci_high": 0.8571428571428571,
|
| 375 |
+
"score_ci_low": 0.14285714285714285,
|
| 376 |
+
"num_of_instances": 7
|
| 377 |
+
},
|
| 378 |
+
"mmlu_pro_engineering": {
|
| 379 |
+
"accuracy": 0.0,
|
| 380 |
+
"accuracy_ci_low": 0.0,
|
| 381 |
+
"accuracy_ci_high": 0.0,
|
| 382 |
+
"score_name": "accuracy",
|
| 383 |
+
"score": 0.0,
|
| 384 |
+
"score_ci_high": 0.0,
|
| 385 |
+
"score_ci_low": 0.0,
|
| 386 |
+
"num_of_instances": 7
|
| 387 |
+
},
|
| 388 |
+
"mmlu_pro_health": {
|
| 389 |
+
"accuracy": 0.5714285714285714,
|
| 390 |
+
"accuracy_ci_low": 0.14285714285714285,
|
| 391 |
+
"accuracy_ci_high": 0.8571428571428571,
|
| 392 |
+
"score_name": "accuracy",
|
| 393 |
+
"score": 0.5714285714285714,
|
| 394 |
+
"score_ci_high": 0.8571428571428571,
|
| 395 |
+
"score_ci_low": 0.14285714285714285,
|
| 396 |
+
"num_of_instances": 7
|
| 397 |
+
},
|
| 398 |
+
"mmlu_pro_history": {
|
| 399 |
+
"accuracy": 0.42857142857142855,
|
| 400 |
+
"accuracy_ci_low": 0.14285714285714285,
|
| 401 |
+
"accuracy_ci_high": 0.8571428571428571,
|
| 402 |
+
"score_name": "accuracy",
|
| 403 |
+
"score": 0.42857142857142855,
|
| 404 |
+
"score_ci_high": 0.8571428571428571,
|
| 405 |
+
"score_ci_low": 0.14285714285714285,
|
| 406 |
+
"num_of_instances": 7
|
| 407 |
+
},
|
| 408 |
+
"mmlu_pro_law": {
|
| 409 |
+
"accuracy": 0.7142857142857143,
|
| 410 |
+
"accuracy_ci_low": 0.2857142857142857,
|
| 411 |
+
"accuracy_ci_high": 1.0,
|
| 412 |
+
"score_name": "accuracy",
|
| 413 |
+
"score": 0.7142857142857143,
|
| 414 |
+
"score_ci_high": 1.0,
|
| 415 |
+
"score_ci_low": 0.2857142857142857,
|
| 416 |
+
"num_of_instances": 7
|
| 417 |
+
},
|
| 418 |
+
"mmlu_pro_math": {
|
| 419 |
+
"accuracy": 0.7142857142857143,
|
| 420 |
+
"accuracy_ci_low": 0.2857142857142857,
|
| 421 |
+
"accuracy_ci_high": 1.0,
|
| 422 |
+
"score_name": "accuracy",
|
| 423 |
+
"score": 0.7142857142857143,
|
| 424 |
+
"score_ci_high": 1.0,
|
| 425 |
+
"score_ci_low": 0.2857142857142857,
|
| 426 |
+
"num_of_instances": 7
|
| 427 |
+
},
|
| 428 |
+
"mmlu_pro_other": {
|
| 429 |
+
"accuracy": 0.2857142857142857,
|
| 430 |
+
"accuracy_ci_low": 0.0,
|
| 431 |
+
"accuracy_ci_high": 0.7142857142857143,
|
| 432 |
+
"score_name": "accuracy",
|
| 433 |
+
"score": 0.2857142857142857,
|
| 434 |
+
"score_ci_high": 0.7142857142857143,
|
| 435 |
+
"score_ci_low": 0.0,
|
| 436 |
+
"num_of_instances": 7
|
| 437 |
+
},
|
| 438 |
+
"mmlu_pro_philosophy": {
|
| 439 |
+
"accuracy": 0.7142857142857143,
|
| 440 |
+
"accuracy_ci_low": 0.2857142857142857,
|
| 441 |
+
"accuracy_ci_high": 1.0,
|
| 442 |
+
"score_name": "accuracy",
|
| 443 |
+
"score": 0.7142857142857143,
|
| 444 |
+
"score_ci_high": 1.0,
|
| 445 |
+
"score_ci_low": 0.2857142857142857,
|
| 446 |
+
"num_of_instances": 7
|
| 447 |
+
},
|
| 448 |
+
"mmlu_pro_physics": {
|
| 449 |
+
"accuracy": 0.14285714285714285,
|
| 450 |
+
"accuracy_ci_low": 0.0,
|
| 451 |
+
"accuracy_ci_high": 0.5714285714285714,
|
| 452 |
+
"score_name": "accuracy",
|
| 453 |
+
"score": 0.14285714285714285,
|
| 454 |
+
"score_ci_high": 0.5714285714285714,
|
| 455 |
+
"score_ci_low": 0.0,
|
| 456 |
+
"num_of_instances": 7
|
| 457 |
+
},
|
| 458 |
+
"mmlu_pro_psychology": {
|
| 459 |
+
"accuracy": 0.7142857142857143,
|
| 460 |
+
"accuracy_ci_low": 0.2857142857142857,
|
| 461 |
+
"accuracy_ci_high": 1.0,
|
| 462 |
+
"score_name": "accuracy",
|
| 463 |
+
"score": 0.7142857142857143,
|
| 464 |
+
"score_ci_high": 1.0,
|
| 465 |
+
"score_ci_low": 0.2857142857142857,
|
| 466 |
+
"num_of_instances": 7
|
| 467 |
+
},
|
| 468 |
+
"score": 0.4489795918367347,
|
| 469 |
+
"score_name": "subsets_mean",
|
| 470 |
+
"num_of_instances": 98
|
| 471 |
+
},
|
| 472 |
+
"legal": {
|
| 473 |
+
"legalbench_abercrombie": {
|
| 474 |
+
"f1_macro": 0.33714285714285713,
|
| 475 |
+
"f1_suggestive": 0.42857142857142855,
|
| 476 |
+
"f1_generic": 0.0,
|
| 477 |
+
"f1_fanciful": 0.4,
|
| 478 |
+
"f1_descriptive": 0.2857142857142857,
|
| 479 |
+
"f1_arbitrary": 0.5714285714285714,
|
| 480 |
+
"f1_macro_ci_low": 0.17333333333333334,
|
| 481 |
+
"f1_macro_ci_high": 0.613790394089633,
|
| 482 |
+
"score_name": "f1_micro",
|
| 483 |
+
"score": 0.3888888888888889,
|
| 484 |
+
"score_ci_high": 0.6131385979389818,
|
| 485 |
+
"score_ci_low": 0.17647058823529413,
|
| 486 |
+
"num_of_instances": 20,
|
| 487 |
+
"accuracy": 0.35,
|
| 488 |
+
"accuracy_ci_low": 0.15,
|
| 489 |
+
"accuracy_ci_high": 0.55,
|
| 490 |
+
"f1_micro": 0.3888888888888889,
|
| 491 |
+
"f1_micro_ci_low": 0.17647058823529413,
|
| 492 |
+
"f1_micro_ci_high": 0.6131385979389818
|
| 493 |
+
},
|
| 494 |
+
"legalbench_corporate_lobbying": {
|
| 495 |
+
"f1_macro": 0.7849462365591398,
|
| 496 |
+
"f1_no": 0.9032258064516129,
|
| 497 |
+
"f1_yes": 0.6666666666666666,
|
| 498 |
+
"f1_macro_ci_low": 0.4546419659069133,
|
| 499 |
+
"f1_macro_ci_high": 1.0,
|
| 500 |
+
"score_name": "f1_micro",
|
| 501 |
+
"score": 0.85,
|
| 502 |
+
"score_ci_high": 0.95,
|
| 503 |
+
"score_ci_low": 0.6,
|
| 504 |
+
"num_of_instances": 20,
|
| 505 |
+
"accuracy": 0.85,
|
| 506 |
+
"accuracy_ci_low": 0.6,
|
| 507 |
+
"accuracy_ci_high": 0.95,
|
| 508 |
+
"f1_micro": 0.85,
|
| 509 |
+
"f1_micro_ci_low": 0.6,
|
| 510 |
+
"f1_micro_ci_high": 0.95
|
| 511 |
+
},
|
| 512 |
+
"legalbench_function_of_decision_section": {
|
| 513 |
+
"f1_macro": 0.19593898165326737,
|
| 514 |
+
"f1_conclusion": 0.2857142857142857,
|
| 515 |
+
"f1_issue": 0.36363636363636365,
|
| 516 |
+
"f1_decree": 0.0,
|
| 517 |
+
"f1_facts": 0.2222222222222222,
|
| 518 |
+
"f1_analysis": 0.0,
|
| 519 |
+
"f1_procedural history": 0.5,
|
| 520 |
+
"f1_rule": 0.0,
|
| 521 |
+
"f1_macro_ci_low": 0.07142857142857142,
|
| 522 |
+
"f1_macro_ci_high": 0.3841799360588237,
|
| 523 |
+
"score_name": "f1_micro",
|
| 524 |
+
"score": 0.2564102564102564,
|
| 525 |
+
"score_ci_high": 0.4864864864864865,
|
| 526 |
+
"score_ci_low": 0.10256410256410256,
|
| 527 |
+
"num_of_instances": 20,
|
| 528 |
+
"accuracy": 0.25,
|
| 529 |
+
"accuracy_ci_low": 0.1,
|
| 530 |
+
"accuracy_ci_high": 0.45,
|
| 531 |
+
"f1_micro": 0.2564102564102564,
|
| 532 |
+
"f1_micro_ci_low": 0.10256410256410256,
|
| 533 |
+
"f1_micro_ci_high": 0.4864864864864865
|
| 534 |
+
},
|
| 535 |
+
"legalbench_international_citizenship_questions": {
|
| 536 |
+
"f1_macro": 0.45054945054945056,
|
| 537 |
+
"f1_yes": 0.6153846153846154,
|
| 538 |
+
"f1_no": 0.2857142857142857,
|
| 539 |
+
"f1_macro_ci_low": 0.2857142857142857,
|
| 540 |
+
"f1_macro_ci_high": 0.696969696969697,
|
| 541 |
+
"score_name": "f1_micro",
|
| 542 |
+
"score": 0.5,
|
| 543 |
+
"score_ci_high": 0.7,
|
| 544 |
+
"score_ci_low": 0.25,
|
| 545 |
+
"num_of_instances": 20,
|
| 546 |
+
"accuracy": 0.5,
|
| 547 |
+
"accuracy_ci_low": 0.25,
|
| 548 |
+
"accuracy_ci_high": 0.7,
|
| 549 |
+
"f1_micro": 0.5,
|
| 550 |
+
"f1_micro_ci_low": 0.25,
|
| 551 |
+
"f1_micro_ci_high": 0.7
|
| 552 |
+
},
|
| 553 |
+
"legalbench_proa": {
|
| 554 |
+
"f1_macro": 0.8465473145780051,
|
| 555 |
+
"f1_yes": 0.8235294117647058,
|
| 556 |
+
"f1_no": 0.8695652173913043,
|
| 557 |
+
"f1_macro_ci_low": 0.6142370542301806,
|
| 558 |
+
"f1_macro_ci_high": 0.949874686716792,
|
| 559 |
+
"score_name": "f1_micro",
|
| 560 |
+
"score": 0.85,
|
| 561 |
+
"score_ci_high": 0.95,
|
| 562 |
+
"score_ci_low": 0.65,
|
| 563 |
+
"num_of_instances": 20,
|
| 564 |
+
"accuracy": 0.85,
|
| 565 |
+
"accuracy_ci_low": 0.65,
|
| 566 |
+
"accuracy_ci_high": 0.95,
|
| 567 |
+
"f1_micro": 0.85,
|
| 568 |
+
"f1_micro_ci_low": 0.65,
|
| 569 |
+
"f1_micro_ci_high": 0.95
|
| 570 |
+
},
|
| 571 |
+
"score": 0.569059829059829,
|
| 572 |
+
"score_name": "subsets_mean",
|
| 573 |
+
"num_of_instances": 100
|
| 574 |
+
},
|
| 575 |
+
"news_classification": {
|
| 576 |
+
"20_newsgroups_short": {
|
| 577 |
+
"f1_macro": 0.3944513819513819,
|
| 578 |
+
"f1_cars": 1.0,
|
| 579 |
+
"f1_windows x": 0.0,
|
| 580 |
+
"f1_atheism": 0.0,
|
| 581 |
+
"f1_christianity": 0.0,
|
| 582 |
+
"f1_religion": 0.0,
|
| 583 |
+
"f1_medicine": 0.8571428571428571,
|
| 584 |
+
"f1_for sale": 0.2222222222222222,
|
| 585 |
+
"f1_computer graphics": 0.36363636363636365,
|
| 586 |
+
"f1_microsoft windows": 0.25,
|
| 587 |
+
"f1_middle east": 0.3333333333333333,
|
| 588 |
+
"f1_politics": 0.46153846153846156,
|
| 589 |
+
"f1_motorcycles": 0.4444444444444444,
|
| 590 |
+
"f1_pc hardware": 0.5714285714285714,
|
| 591 |
+
"f1_mac hardware": 0.2857142857142857,
|
| 592 |
+
"f1_electronics": 0.2857142857142857,
|
| 593 |
+
"f1_guns": 0.5,
|
| 594 |
+
"f1_space": 0.5,
|
| 595 |
+
"f1_cryptography": 0.3333333333333333,
|
| 596 |
+
"f1_baseball": 0.9090909090909091,
|
| 597 |
+
"f1_hockey": 0.5714285714285714,
|
| 598 |
+
"f1_macro_ci_low": 0.3250051046310291,
|
| 599 |
+
"f1_macro_ci_high": 0.4955223374328094,
|
| 600 |
+
"score_name": "f1_micro",
|
| 601 |
+
"score": 0.42162162162162165,
|
| 602 |
+
"score_ci_high": 0.521953556122739,
|
| 603 |
+
"score_ci_low": 0.31511577882586567,
|
| 604 |
+
"num_of_instances": 100,
|
| 605 |
+
"accuracy": 0.39,
|
| 606 |
+
"accuracy_ci_low": 0.3,
|
| 607 |
+
"accuracy_ci_high": 0.4918126232007319,
|
| 608 |
+
"f1_micro": 0.42162162162162165,
|
| 609 |
+
"f1_micro_ci_low": 0.31511577882586567,
|
| 610 |
+
"f1_micro_ci_high": 0.521953556122739
|
| 611 |
+
},
|
| 612 |
+
"score": 0.42162162162162165,
|
| 613 |
+
"score_name": "subsets_mean",
|
| 614 |
+
"num_of_instances": 100
|
| 615 |
+
},
|
| 616 |
+
"product_help": {
|
| 617 |
+
"cfpb_product_2023": {
|
| 618 |
+
"f1_macro": 0.5114463197955563,
|
| 619 |
+
"f1_credit reporting or credit repair services or other personal consumer reports": 0.8396946564885496,
|
| 620 |
+
"f1_debt collection": 0.6,
|
| 621 |
+
"f1_payday loan or title loan or personal loan": 0.0,
|
| 622 |
+
"f1_credit card or prepaid card": 0.2857142857142857,
|
| 623 |
+
"f1_student loan": 0.8888888888888888,
|
| 624 |
+
"f1_checking or savings account": 0.7272727272727273,
|
| 625 |
+
"f1_mortgage": 0.75,
|
| 626 |
+
"f1_money transfer or virtual currency or money service": 0.0,
|
| 627 |
+
"f1_macro_ci_low": 0.33282929006231693,
|
| 628 |
+
"f1_macro_ci_high": 0.6228547419869556,
|
| 629 |
+
"score_name": "f1_micro",
|
| 630 |
+
"score": 0.7724867724867724,
|
| 631 |
+
"score_ci_high": 0.845360824742268,
|
| 632 |
+
"score_ci_low": 0.6774193548387096,
|
| 633 |
+
"num_of_instances": 100,
|
| 634 |
+
"accuracy": 0.73,
|
| 635 |
+
"accuracy_ci_low": 0.63,
|
| 636 |
+
"accuracy_ci_high": 0.81,
|
| 637 |
+
"f1_micro": 0.7724867724867724,
|
| 638 |
+
"f1_micro_ci_low": 0.6774193548387096,
|
| 639 |
+
"f1_micro_ci_high": 0.845360824742268
|
| 640 |
+
},
|
| 641 |
+
"cfpb_product_watsonx": {
|
| 642 |
+
"f1_macro": 0.5816666666666667,
|
| 643 |
+
"f1_mortgages and loans": 0.7,
|
| 644 |
+
"f1_credit card": 0.5,
|
| 645 |
+
"f1_credit reporting": 0.6666666666666666,
|
| 646 |
+
"f1_retail banking": 0.375,
|
| 647 |
+
"f1_debt collection": 0.6666666666666666,
|
| 648 |
+
"f1_macro_ci_low": 0.45364221453455666,
|
| 649 |
+
"f1_macro_ci_high": 0.7245142399435678,
|
| 650 |
+
"score_name": "f1_micro",
|
| 651 |
+
"score": 0.5979381443298969,
|
| 652 |
+
"score_ci_high": 0.7216494845360825,
|
| 653 |
+
"score_ci_low": 0.4489795918367347,
|
| 654 |
+
"num_of_instances": 50,
|
| 655 |
+
"accuracy": 0.58,
|
| 656 |
+
"accuracy_ci_low": 0.44,
|
| 657 |
+
"accuracy_ci_high": 0.7,
|
| 658 |
+
"f1_micro": 0.5979381443298969,
|
| 659 |
+
"f1_micro_ci_low": 0.4489795918367347,
|
| 660 |
+
"f1_micro_ci_high": 0.7216494845360825
|
| 661 |
+
},
|
| 662 |
+
"score": 0.6852124584083347,
|
| 663 |
+
"score_name": "subsets_mean",
|
| 664 |
+
"num_of_instances": 150
|
| 665 |
+
},
|
| 666 |
+
"qa_finance": {
|
| 667 |
+
"fin_qa": {
|
| 668 |
+
"num_of_instances": 100,
|
| 669 |
+
"execution_accuracy": 0.24,
|
| 670 |
+
"program_accuracy": 0.26,
|
| 671 |
+
"score": 0.26,
|
| 672 |
+
"score_name": "program_accuracy",
|
| 673 |
+
"execution_accuracy_ci_low": 0.16,
|
| 674 |
+
"execution_accuracy_ci_high": 0.33,
|
| 675 |
+
"program_accuracy_ci_low": 0.18,
|
| 676 |
+
"program_accuracy_ci_high": 0.35,
|
| 677 |
+
"score_ci_low": 0.18,
|
| 678 |
+
"score_ci_high": 0.35
|
| 679 |
+
},
|
| 680 |
+
"score": 0.26,
|
| 681 |
+
"score_name": "subsets_mean",
|
| 682 |
+
"num_of_instances": 100
|
| 683 |
+
},
|
| 684 |
+
"rag_general": {
|
| 685 |
+
"rag_response_generation_clapnq": {
|
| 686 |
+
"precision": 0.5483530944738381,
|
| 687 |
+
"recall": 0.5554279823377488,
|
| 688 |
+
"f1": 0.5087588370039454,
|
| 689 |
+
"precision_ci_low": 0.5059194304671544,
|
| 690 |
+
"precision_ci_high": 0.586130205045163,
|
| 691 |
+
"recall_ci_low": 0.5103407537092706,
|
| 692 |
+
"recall_ci_high": 0.5969208798184268,
|
| 693 |
+
"f1_ci_low": 0.4762966268521691,
|
| 694 |
+
"f1_ci_high": 0.5419921191805944,
|
| 695 |
+
"score_name": "f1",
|
| 696 |
+
"score": 0.5087588370039454,
|
| 697 |
+
"score_ci_high": 0.5419921191805944,
|
| 698 |
+
"score_ci_low": 0.4762966268521691,
|
| 699 |
+
"num_of_instances": 100,
|
| 700 |
+
"correctness_f1_bert_score.deberta_large_mnli": 0.707559947669506,
|
| 701 |
+
"correctness_recall_bert_score.deberta_large_mnli": 0.7081431838870048,
|
| 702 |
+
"correctness_precision_bert_score.deberta_large_mnli": 0.7173768043518066,
|
| 703 |
+
"faithfullness_f1_token_overlap": 0.3520099489303385,
|
| 704 |
+
"faithfullness_recall_token_overlap": 0.25507157752637033,
|
| 705 |
+
"faithfullness_precision_token_overlap": 0.7306707307941289,
|
| 706 |
+
"correctness_f1_token_overlap": 0.5087588370039454,
|
| 707 |
+
"correctness_recall_token_overlap": 0.5554279823377488,
|
| 708 |
+
"correctness_precision_token_overlap": 0.5483530944738381
|
| 709 |
+
},
|
| 710 |
+
"score": 0.5087588370039454,
|
| 711 |
+
"score_name": "subsets_mean",
|
| 712 |
+
"num_of_instances": 100
|
| 713 |
+
},
|
| 714 |
+
"reasoning": {
|
| 715 |
+
"hellaswag": {
|
| 716 |
+
"accuracy": 0.51,
|
| 717 |
+
"accuracy_ci_low": 0.4,
|
| 718 |
+
"accuracy_ci_high": 0.6,
|
| 719 |
+
"score_name": "accuracy",
|
| 720 |
+
"score": 0.51,
|
| 721 |
+
"score_ci_high": 0.6,
|
| 722 |
+
"score_ci_low": 0.4,
|
| 723 |
+
"num_of_instances": 100
|
| 724 |
+
},
|
| 725 |
+
"openbook_qa": {
|
| 726 |
+
"accuracy": 0.87,
|
| 727 |
+
"accuracy_ci_low": 0.79,
|
| 728 |
+
"accuracy_ci_high": 0.93,
|
| 729 |
+
"score_name": "accuracy",
|
| 730 |
+
"score": 0.87,
|
| 731 |
+
"score_ci_high": 0.93,
|
| 732 |
+
"score_ci_low": 0.79,
|
| 733 |
+
"num_of_instances": 100
|
| 734 |
+
},
|
| 735 |
+
"score": 0.69,
|
| 736 |
+
"score_name": "subsets_mean",
|
| 737 |
+
"num_of_instances": 200
|
| 738 |
+
},
|
| 739 |
+
"safety": {
|
| 740 |
+
"attaq_500": {
|
| 741 |
+
"severity_critical": 0.0,
|
| 742 |
+
"severity_high": 0.0,
|
| 743 |
+
"severity_medium": 1.0,
|
| 744 |
+
"severity_low": 99.0,
|
| 745 |
+
"category_harmful_info": 0.8783717133356111,
|
| 746 |
+
"category_deception": 0.8950554424558174,
|
| 747 |
+
"category_explicit_content": 0.8705809025300874,
|
| 748 |
+
"category_violence": 0.8501951797732284,
|
| 749 |
+
"category_discrimination": 0.8293574092880128,
|
| 750 |
+
"category_substance_abuse": 0.888910385966301,
|
| 751 |
+
"category_pii": 0.8837605582343208,
|
| 752 |
+
"safety": 0.8667457452985563,
|
| 753 |
+
"safety_ci_low": 0.848787173973939,
|
| 754 |
+
"safety_ci_high": 0.8831856051728106,
|
| 755 |
+
"score_name": "safety",
|
| 756 |
+
"score": 0.8667457452985563,
|
| 757 |
+
"score_ci_high": 0.8831856051728106,
|
| 758 |
+
"score_ci_low": 0.848787173973939,
|
| 759 |
+
"num_of_instances": 100
|
| 760 |
+
},
|
| 761 |
+
"score": 0.8667457452985563,
|
| 762 |
+
"score_name": "subsets_mean",
|
| 763 |
+
"num_of_instances": 100
|
| 764 |
+
},
|
| 765 |
+
"summarization": {
|
| 766 |
+
"billsum_document_filtered_to_6000_chars": {
|
| 767 |
+
"num_of_instances": 100,
|
| 768 |
+
"rougeLsum": 0.3223733838472141,
|
| 769 |
+
"rouge2": 0.16381533114294386,
|
| 770 |
+
"rouge1": 0.38044015102427337,
|
| 771 |
+
"rougeL": 0.25853052606402777,
|
| 772 |
+
"score": 0.25853052606402777,
|
| 773 |
+
"score_name": "rougeL",
|
| 774 |
+
"rougeLsum_ci_low": 0.3017771809602484,
|
| 775 |
+
"rougeLsum_ci_high": 0.34134015771556503,
|
| 776 |
+
"rouge2_ci_low": 0.14895504468906468,
|
| 777 |
+
"rouge2_ci_high": 0.1764309652967554,
|
| 778 |
+
"rouge1_ci_low": 0.3558190631222209,
|
| 779 |
+
"rouge1_ci_high": 0.3991166452082928,
|
| 780 |
+
"rougeL_ci_low": 0.24136665350639253,
|
| 781 |
+
"rougeL_ci_high": 0.2728771499617261,
|
| 782 |
+
"score_ci_low": 0.24136665350639253,
|
| 783 |
+
"score_ci_high": 0.2728771499617261
|
| 784 |
+
},
|
| 785 |
+
"tldr_document_filtered_to_6000_chars": {
|
| 786 |
+
"num_of_instances": 100,
|
| 787 |
+
"rougeLsum": 0.08773717788599067,
|
| 788 |
+
"rouge2": 0.013999577794814255,
|
| 789 |
+
"rouge1": 0.10722757537209487,
|
| 790 |
+
"rougeL": 0.08028090646112783,
|
| 791 |
+
"score": 0.08028090646112783,
|
| 792 |
+
"score_name": "rougeL",
|
| 793 |
+
"rougeLsum_ci_low": 0.0761926584169636,
|
| 794 |
+
"rougeLsum_ci_high": 0.09991471571884306,
|
| 795 |
+
"rouge2_ci_low": 0.009830037235289218,
|
| 796 |
+
"rouge2_ci_high": 0.020298396822460246,
|
| 797 |
+
"rouge1_ci_low": 0.09194762041845511,
|
| 798 |
+
"rouge1_ci_high": 0.1237270589822993,
|
| 799 |
+
"rougeL_ci_low": 0.06967530823212387,
|
| 800 |
+
"rougeL_ci_high": 0.09119370435892198,
|
| 801 |
+
"score_ci_low": 0.06967530823212387,
|
| 802 |
+
"score_ci_high": 0.09119370435892198
|
| 803 |
+
},
|
| 804 |
+
"score": 0.1694057162625778,
|
| 805 |
+
"score_name": "subsets_mean",
|
| 806 |
+
"num_of_instances": 200
|
| 807 |
+
},
|
| 808 |
+
"translation": {
|
| 809 |
+
"mt_flores_101_ara_eng": {
|
| 810 |
+
"num_of_instances": 6,
|
| 811 |
+
"counts": [
|
| 812 |
+
153,
|
| 813 |
+
107,
|
| 814 |
+
81,
|
| 815 |
+
65
|
| 816 |
+
],
|
| 817 |
+
"totals": [
|
| 818 |
+
216,
|
| 819 |
+
210,
|
| 820 |
+
204,
|
| 821 |
+
198
|
| 822 |
+
],
|
| 823 |
+
"precisions": [
|
| 824 |
+
0.7083333333333333,
|
| 825 |
+
0.5095238095238095,
|
| 826 |
+
0.39705882352941174,
|
| 827 |
+
0.3282828282828283
|
| 828 |
+
],
|
| 829 |
+
"bp": 1.0,
|
| 830 |
+
"sys_len": 216,
|
| 831 |
+
"ref_len": 208,
|
| 832 |
+
"sacrebleu": 0.4657215080894149,
|
| 833 |
+
"score": 0.4657215080894149,
|
| 834 |
+
"score_name": "sacrebleu",
|
| 835 |
+
"score_ci_low": 0.26113318051473333,
|
| 836 |
+
"score_ci_high": 0.554322605325799,
|
| 837 |
+
"sacrebleu_ci_low": 0.26113318051473333,
|
| 838 |
+
"sacrebleu_ci_high": 0.554322605325799
|
| 839 |
+
},
|
| 840 |
+
"mt_flores_101_deu_eng": {
|
| 841 |
+
"num_of_instances": 6,
|
| 842 |
+
"counts": [
|
| 843 |
+
140,
|
| 844 |
+
88,
|
| 845 |
+
54,
|
| 846 |
+
39
|
| 847 |
+
],
|
| 848 |
+
"totals": [
|
| 849 |
+
219,
|
| 850 |
+
213,
|
| 851 |
+
207,
|
| 852 |
+
201
|
| 853 |
+
],
|
| 854 |
+
"precisions": [
|
| 855 |
+
0.639269406392694,
|
| 856 |
+
0.41314553990610325,
|
| 857 |
+
0.2608695652173913,
|
| 858 |
+
0.19402985074626866
|
| 859 |
+
],
|
| 860 |
+
"bp": 1.0,
|
| 861 |
+
"sys_len": 219,
|
| 862 |
+
"ref_len": 208,
|
| 863 |
+
"sacrebleu": 0.34003195966759087,
|
| 864 |
+
"score": 0.34003195966759087,
|
| 865 |
+
"score_name": "sacrebleu",
|
| 866 |
+
"score_ci_low": 0.23265738389765314,
|
| 867 |
+
"score_ci_high": 0.5005821053278499,
|
| 868 |
+
"sacrebleu_ci_low": 0.23265738389765314,
|
| 869 |
+
"sacrebleu_ci_high": 0.5005821053278499
|
| 870 |
+
},
|
| 871 |
+
"mt_flores_101_eng_ara": {
|
| 872 |
+
"num_of_instances": 6,
|
| 873 |
+
"counts": [
|
| 874 |
+
112,
|
| 875 |
+
59,
|
| 876 |
+
32,
|
| 877 |
+
18
|
| 878 |
+
],
|
| 879 |
+
"totals": [
|
| 880 |
+
203,
|
| 881 |
+
197,
|
| 882 |
+
191,
|
| 883 |
+
185
|
| 884 |
+
],
|
| 885 |
+
"precisions": [
|
| 886 |
+
0.5517241379310345,
|
| 887 |
+
0.29949238578680204,
|
| 888 |
+
0.16753926701570682,
|
| 889 |
+
0.0972972972972973
|
| 890 |
+
],
|
| 891 |
+
"bp": 0.9708758757257812,
|
| 892 |
+
"sys_len": 203,
|
| 893 |
+
"ref_len": 209,
|
| 894 |
+
"sacrebleu": 0.2211795649032297,
|
| 895 |
+
"score": 0.2211795649032297,
|
| 896 |
+
"score_name": "sacrebleu",
|
| 897 |
+
"score_ci_low": 0.16372679305283608,
|
| 898 |
+
"score_ci_high": 0.29635725852266814,
|
| 899 |
+
"sacrebleu_ci_low": 0.16372679305283608,
|
| 900 |
+
"sacrebleu_ci_high": 0.29635725852266814
|
| 901 |
+
},
|
| 902 |
+
"mt_flores_101_eng_deu": {
|
| 903 |
+
"num_of_instances": 6,
|
| 904 |
+
"counts": [
|
| 905 |
+
138,
|
| 906 |
+
84,
|
| 907 |
+
55,
|
| 908 |
+
36
|
| 909 |
+
],
|
| 910 |
+
"totals": [
|
| 911 |
+
220,
|
| 912 |
+
214,
|
| 913 |
+
208,
|
| 914 |
+
202
|
| 915 |
+
],
|
| 916 |
+
"precisions": [
|
| 917 |
+
0.6272727272727273,
|
| 918 |
+
0.3925233644859813,
|
| 919 |
+
0.2644230769230769,
|
| 920 |
+
0.17821782178217824
|
| 921 |
+
],
|
| 922 |
+
"bp": 1.0,
|
| 923 |
+
"sys_len": 220,
|
| 924 |
+
"ref_len": 216,
|
| 925 |
+
"sacrebleu": 0.3282034190837251,
|
| 926 |
+
"score": 0.3282034190837251,
|
| 927 |
+
"score_name": "sacrebleu",
|
| 928 |
+
"score_ci_low": 0.18070782395940485,
|
| 929 |
+
"score_ci_high": 0.46287226401772935,
|
| 930 |
+
"sacrebleu_ci_low": 0.18070782395940485,
|
| 931 |
+
"sacrebleu_ci_high": 0.46287226401772935
|
| 932 |
+
},
|
| 933 |
+
"mt_flores_101_eng_fra": {
|
| 934 |
+
"num_of_instances": 6,
|
| 935 |
+
"counts": [
|
| 936 |
+
189,
|
| 937 |
+
149,
|
| 938 |
+
120,
|
| 939 |
+
99
|
| 940 |
+
],
|
| 941 |
+
"totals": [
|
| 942 |
+
233,
|
| 943 |
+
227,
|
| 944 |
+
221,
|
| 945 |
+
215
|
| 946 |
+
],
|
| 947 |
+
"precisions": [
|
| 948 |
+
0.8111587982832619,
|
| 949 |
+
0.6563876651982379,
|
| 950 |
+
0.5429864253393665,
|
| 951 |
+
0.4604651162790697
|
| 952 |
+
],
|
| 953 |
+
"bp": 0.9914530437067961,
|
| 954 |
+
"sys_len": 233,
|
| 955 |
+
"ref_len": 235,
|
| 956 |
+
"sacrebleu": 0.5988735753373767,
|
| 957 |
+
"score": 0.5988735753373767,
|
| 958 |
+
"score_name": "sacrebleu",
|
| 959 |
+
"score_ci_low": 0.4819198427452047,
|
| 960 |
+
"score_ci_high": 0.7080462528698704,
|
| 961 |
+
"sacrebleu_ci_low": 0.4819198427452047,
|
| 962 |
+
"sacrebleu_ci_high": 0.7080462528698704
|
| 963 |
+
},
|
| 964 |
+
"mt_flores_101_eng_kor": {
|
| 965 |
+
"num_of_instances": 6,
|
| 966 |
+
"counts": [
|
| 967 |
+
155,
|
| 968 |
+
85,
|
| 969 |
+
52,
|
| 970 |
+
30
|
| 971 |
+
],
|
| 972 |
+
"totals": [
|
| 973 |
+
276,
|
| 974 |
+
270,
|
| 975 |
+
264,
|
| 976 |
+
258
|
| 977 |
+
],
|
| 978 |
+
"precisions": [
|
| 979 |
+
0.5615942028985507,
|
| 980 |
+
0.3148148148148148,
|
| 981 |
+
0.19696969696969696,
|
| 982 |
+
0.11627906976744186
|
| 983 |
+
],
|
| 984 |
+
"bp": 1.0,
|
| 985 |
+
"sys_len": 276,
|
| 986 |
+
"ref_len": 249,
|
| 987 |
+
"sacrebleu": 0.25225784761095216,
|
| 988 |
+
"score": 0.25225784761095216,
|
| 989 |
+
"score_name": "sacrebleu",
|
| 990 |
+
"score_ci_low": 0.1938128536412581,
|
| 991 |
+
"score_ci_high": 0.3424326666483488,
|
| 992 |
+
"sacrebleu_ci_low": 0.1938128536412581,
|
| 993 |
+
"sacrebleu_ci_high": 0.3424326666483488
|
| 994 |
+
},
|
| 995 |
+
"mt_flores_101_eng_por": {
|
| 996 |
+
"num_of_instances": 6,
|
| 997 |
+
"counts": [
|
| 998 |
+
181,
|
| 999 |
+
139,
|
| 1000 |
+
115,
|
| 1001 |
+
95
|
| 1002 |
+
],
|
| 1003 |
+
"totals": [
|
| 1004 |
+
217,
|
| 1005 |
+
211,
|
| 1006 |
+
205,
|
| 1007 |
+
199
|
| 1008 |
+
],
|
| 1009 |
+
"precisions": [
|
| 1010 |
+
0.8341013824884793,
|
| 1011 |
+
0.6587677725118484,
|
| 1012 |
+
0.5609756097560975,
|
| 1013 |
+
0.4773869346733668
|
| 1014 |
+
],
|
| 1015 |
+
"bp": 0.977221952990032,
|
| 1016 |
+
"sys_len": 217,
|
| 1017 |
+
"ref_len": 222,
|
| 1018 |
+
"sacrebleu": 0.6052497771952972,
|
| 1019 |
+
"score": 0.6052497771952972,
|
| 1020 |
+
"score_name": "sacrebleu",
|
| 1021 |
+
"score_ci_low": 0.5137095133218235,
|
| 1022 |
+
"score_ci_high": 0.7430807912893931,
|
| 1023 |
+
"sacrebleu_ci_low": 0.5137095133218235,
|
| 1024 |
+
"sacrebleu_ci_high": 0.7430807912893931
|
| 1025 |
+
},
|
| 1026 |
+
"mt_flores_101_eng_ron": {
|
| 1027 |
+
"num_of_instances": 6,
|
| 1028 |
+
"counts": [
|
| 1029 |
+
161,
|
| 1030 |
+
114,
|
| 1031 |
+
82,
|
| 1032 |
+
64
|
| 1033 |
+
],
|
| 1034 |
+
"totals": [
|
| 1035 |
+
226,
|
| 1036 |
+
220,
|
| 1037 |
+
214,
|
| 1038 |
+
208
|
| 1039 |
+
],
|
| 1040 |
+
"precisions": [
|
| 1041 |
+
0.7123893805309734,
|
| 1042 |
+
0.5181818181818182,
|
| 1043 |
+
0.383177570093458,
|
| 1044 |
+
0.3076923076923077
|
| 1045 |
+
],
|
| 1046 |
+
"bp": 0.9824565942999044,
|
| 1047 |
+
"sys_len": 226,
|
| 1048 |
+
"ref_len": 230,
|
| 1049 |
+
"sacrebleu": 0.4487375922561012,
|
| 1050 |
+
"score": 0.4487375922561012,
|
| 1051 |
+
"score_name": "sacrebleu",
|
| 1052 |
+
"score_ci_low": 0.3680871645957313,
|
| 1053 |
+
"score_ci_high": 0.5551991202375558,
|
| 1054 |
+
"sacrebleu_ci_low": 0.3680871645957313,
|
| 1055 |
+
"sacrebleu_ci_high": 0.5551991202375558
|
| 1056 |
+
},
|
| 1057 |
+
"mt_flores_101_eng_spa": {
|
| 1058 |
+
"num_of_instances": 6,
|
| 1059 |
+
"counts": [
|
| 1060 |
+
160,
|
| 1061 |
+
97,
|
| 1062 |
+
61,
|
| 1063 |
+
40
|
| 1064 |
+
],
|
| 1065 |
+
"totals": [
|
| 1066 |
+
232,
|
| 1067 |
+
226,
|
| 1068 |
+
220,
|
| 1069 |
+
214
|
| 1070 |
+
],
|
| 1071 |
+
"precisions": [
|
| 1072 |
+
0.6896551724137931,
|
| 1073 |
+
0.4292035398230089,
|
| 1074 |
+
0.2772727272727273,
|
| 1075 |
+
0.18691588785046728
|
| 1076 |
+
],
|
| 1077 |
+
"bp": 0.9536926844755759,
|
| 1078 |
+
"sys_len": 232,
|
| 1079 |
+
"ref_len": 243,
|
| 1080 |
+
"sacrebleu": 0.3356376081723427,
|
| 1081 |
+
"score": 0.3356376081723427,
|
| 1082 |
+
"score_name": "sacrebleu",
|
| 1083 |
+
"score_ci_low": 0.27756379712473256,
|
| 1084 |
+
"score_ci_high": 0.4039979777599405,
|
| 1085 |
+
"sacrebleu_ci_low": 0.27756379712473256,
|
| 1086 |
+
"sacrebleu_ci_high": 0.4039979777599405
|
| 1087 |
+
},
|
| 1088 |
+
"mt_flores_101_fra_eng": {
|
| 1089 |
+
"num_of_instances": 6,
|
| 1090 |
+
"counts": [
|
| 1091 |
+
165,
|
| 1092 |
+
124,
|
| 1093 |
+
98,
|
| 1094 |
+
79
|
| 1095 |
+
],
|
| 1096 |
+
"totals": [
|
| 1097 |
+
220,
|
| 1098 |
+
214,
|
| 1099 |
+
208,
|
| 1100 |
+
202
|
| 1101 |
+
],
|
| 1102 |
+
"precisions": [
|
| 1103 |
+
0.75,
|
| 1104 |
+
0.5794392523364487,
|
| 1105 |
+
0.47115384615384615,
|
| 1106 |
+
0.3910891089108911
|
| 1107 |
+
],
|
| 1108 |
+
"bp": 1.0,
|
| 1109 |
+
"sys_len": 220,
|
| 1110 |
+
"ref_len": 208,
|
| 1111 |
+
"sacrebleu": 0.5319574670672091,
|
| 1112 |
+
"score": 0.5319574670672091,
|
| 1113 |
+
"score_name": "sacrebleu",
|
| 1114 |
+
"score_ci_low": 0.39796202118207913,
|
| 1115 |
+
"score_ci_high": 0.6612896148004259,
|
| 1116 |
+
"sacrebleu_ci_low": 0.39796202118207913,
|
| 1117 |
+
"sacrebleu_ci_high": 0.6612896148004259
|
| 1118 |
+
},
|
| 1119 |
+
"mt_flores_101_jpn_eng": {
|
| 1120 |
+
"num_of_instances": 6,
|
| 1121 |
+
"counts": [
|
| 1122 |
+
135,
|
| 1123 |
+
70,
|
| 1124 |
+
37,
|
| 1125 |
+
21
|
| 1126 |
+
],
|
| 1127 |
+
"totals": [
|
| 1128 |
+
216,
|
| 1129 |
+
210,
|
| 1130 |
+
204,
|
| 1131 |
+
198
|
| 1132 |
+
],
|
| 1133 |
+
"precisions": [
|
| 1134 |
+
0.625,
|
| 1135 |
+
0.33333333333333337,
|
| 1136 |
+
0.18137254901960784,
|
| 1137 |
+
0.10606060606060605
|
| 1138 |
+
],
|
| 1139 |
+
"bp": 1.0,
|
| 1140 |
+
"sys_len": 216,
|
| 1141 |
+
"ref_len": 208,
|
| 1142 |
+
"sacrebleu": 0.2516060651765726,
|
| 1143 |
+
"score": 0.2516060651765726,
|
| 1144 |
+
"score_name": "sacrebleu",
|
| 1145 |
+
"score_ci_low": 0.12146602398816096,
|
| 1146 |
+
"score_ci_high": 0.3230294156821773,
|
| 1147 |
+
"sacrebleu_ci_low": 0.12146602398816096,
|
| 1148 |
+
"sacrebleu_ci_high": 0.3230294156821773
|
| 1149 |
+
},
|
| 1150 |
+
"mt_flores_101_kor_eng": {
|
| 1151 |
+
"num_of_instances": 6,
|
| 1152 |
+
"counts": [
|
| 1153 |
+
136,
|
| 1154 |
+
78,
|
| 1155 |
+
50,
|
| 1156 |
+
35
|
| 1157 |
+
],
|
| 1158 |
+
"totals": [
|
| 1159 |
+
212,
|
| 1160 |
+
206,
|
| 1161 |
+
200,
|
| 1162 |
+
194
|
| 1163 |
+
],
|
| 1164 |
+
"precisions": [
|
| 1165 |
+
0.6415094339622641,
|
| 1166 |
+
0.3786407766990291,
|
| 1167 |
+
0.25,
|
| 1168 |
+
0.18041237113402062
|
| 1169 |
+
],
|
| 1170 |
+
"bp": 1.0,
|
| 1171 |
+
"sys_len": 212,
|
| 1172 |
+
"ref_len": 208,
|
| 1173 |
+
"sacrebleu": 0.32352599996619197,
|
| 1174 |
+
"score": 0.32352599996619197,
|
| 1175 |
+
"score_name": "sacrebleu",
|
| 1176 |
+
"score_ci_low": 0.20829290856129712,
|
| 1177 |
+
"score_ci_high": 0.45329328917319234,
|
| 1178 |
+
"sacrebleu_ci_low": 0.20829290856129712,
|
| 1179 |
+
"sacrebleu_ci_high": 0.45329328917319234
|
| 1180 |
+
},
|
| 1181 |
+
"mt_flores_101_por_eng": {
|
| 1182 |
+
"num_of_instances": 6,
|
| 1183 |
+
"counts": [
|
| 1184 |
+
166,
|
| 1185 |
+
128,
|
| 1186 |
+
98,
|
| 1187 |
+
77
|
| 1188 |
+
],
|
| 1189 |
+
"totals": [
|
| 1190 |
+
212,
|
| 1191 |
+
206,
|
| 1192 |
+
200,
|
| 1193 |
+
194
|
| 1194 |
+
],
|
| 1195 |
+
"precisions": [
|
| 1196 |
+
0.7830188679245284,
|
| 1197 |
+
0.6213592233009709,
|
| 1198 |
+
0.49,
|
| 1199 |
+
0.3969072164948454
|
| 1200 |
+
],
|
| 1201 |
+
"bp": 1.0,
|
| 1202 |
+
"sys_len": 212,
|
| 1203 |
+
"ref_len": 208,
|
| 1204 |
+
"sacrebleu": 0.5546257294515591,
|
| 1205 |
+
"score": 0.5546257294515591,
|
| 1206 |
+
"score_name": "sacrebleu",
|
| 1207 |
+
"score_ci_low": 0.4041829047396742,
|
| 1208 |
+
"score_ci_high": 0.6689609866438883,
|
| 1209 |
+
"sacrebleu_ci_low": 0.4041829047396742,
|
| 1210 |
+
"sacrebleu_ci_high": 0.6689609866438883
|
| 1211 |
+
},
|
| 1212 |
+
"mt_flores_101_ron_eng": {
|
| 1213 |
+
"num_of_instances": 6,
|
| 1214 |
+
"counts": [
|
| 1215 |
+
155,
|
| 1216 |
+
106,
|
| 1217 |
+
74,
|
| 1218 |
+
52
|
| 1219 |
+
],
|
| 1220 |
+
"totals": [
|
| 1221 |
+
224,
|
| 1222 |
+
218,
|
| 1223 |
+
212,
|
| 1224 |
+
206
|
| 1225 |
+
],
|
| 1226 |
+
"precisions": [
|
| 1227 |
+
0.6919642857142857,
|
| 1228 |
+
0.48623853211009177,
|
| 1229 |
+
0.34905660377358494,
|
| 1230 |
+
0.2524271844660194
|
| 1231 |
+
],
|
| 1232 |
+
"bp": 1.0,
|
| 1233 |
+
"sys_len": 224,
|
| 1234 |
+
"ref_len": 208,
|
| 1235 |
+
"sacrebleu": 0.41494569039959667,
|
| 1236 |
+
"score": 0.41494569039959667,
|
| 1237 |
+
"score_name": "sacrebleu",
|
| 1238 |
+
"score_ci_low": 0.36976663169974777,
|
| 1239 |
+
"score_ci_high": 0.5018256593904208,
|
| 1240 |
+
"sacrebleu_ci_low": 0.36976663169974777,
|
| 1241 |
+
"sacrebleu_ci_high": 0.5018256593904208
|
| 1242 |
+
},
|
| 1243 |
+
"mt_flores_101_spa_eng": {
|
| 1244 |
+
"num_of_instances": 6,
|
| 1245 |
+
"counts": [
|
| 1246 |
+
150,
|
| 1247 |
+
97,
|
| 1248 |
+
63,
|
| 1249 |
+
43
|
| 1250 |
+
],
|
| 1251 |
+
"totals": [
|
| 1252 |
+
223,
|
| 1253 |
+
217,
|
| 1254 |
+
211,
|
| 1255 |
+
205
|
| 1256 |
+
],
|
| 1257 |
+
"precisions": [
|
| 1258 |
+
0.6726457399103138,
|
| 1259 |
+
0.4470046082949309,
|
| 1260 |
+
0.2985781990521327,
|
| 1261 |
+
0.20975609756097563
|
| 1262 |
+
],
|
| 1263 |
+
"bp": 1.0,
|
| 1264 |
+
"sys_len": 223,
|
| 1265 |
+
"ref_len": 208,
|
| 1266 |
+
"sacrebleu": 0.37043991107495844,
|
| 1267 |
+
"score": 0.37043991107495844,
|
| 1268 |
+
"score_name": "sacrebleu",
|
| 1269 |
+
"score_ci_low": 0.341504023583382,
|
| 1270 |
+
"score_ci_high": 0.41393253157059545,
|
| 1271 |
+
"sacrebleu_ci_low": 0.341504023583382,
|
| 1272 |
+
"sacrebleu_ci_high": 0.41393253157059545
|
| 1273 |
+
},
|
| 1274 |
+
"score": 0.4028662476968079,
|
| 1275 |
+
"score_name": "subsets_mean",
|
| 1276 |
+
"num_of_instances": 90
|
| 1277 |
+
},
|
| 1278 |
+
"score": 0.5698160000855291,
|
| 1279 |
+
"score_name": "subsets_mean",
|
| 1280 |
+
"num_of_instances": 1537
|
| 1281 |
+
}
|
| 1282 |
+
}
|