more gpt results
Browse filesSigned-off-by: Jonathan Bnayahu <[email protected]>
results/bluebench/2025-08-03T09-40-01_evaluation_results.json
ADDED
|
@@ -0,0 +1,1282 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"environment_info": {
|
| 3 |
+
"timestamp_utc": "2025-08-03T13:39:57.204417Z",
|
| 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-4o-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-4o-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": 0.8888888888888888,
|
| 180 |
+
"accuracy_ci_low": 0.4444444444444444,
|
| 181 |
+
"accuracy_ci_high": 1.0,
|
| 182 |
+
"score_name": "accuracy",
|
| 183 |
+
"score": 0.8888888888888888,
|
| 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": 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": 1.0,
|
| 260 |
+
"accuracy_ci_low": 1.0,
|
| 261 |
+
"accuracy_ci_high": 1.0,
|
| 262 |
+
"score_name": "accuracy",
|
| 263 |
+
"score": 1.0,
|
| 264 |
+
"score_ci_high": 1.0,
|
| 265 |
+
"score_ci_low": 1.0,
|
| 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": 0.5555555555555556,
|
| 281 |
+
"accuracy_ci_high": 1.0,
|
| 282 |
+
"score_name": "accuracy",
|
| 283 |
+
"score": 0.8888888888888888,
|
| 284 |
+
"score_ci_high": 1.0,
|
| 285 |
+
"score_ci_low": 0.5555555555555556,
|
| 286 |
+
"num_of_instances": 9
|
| 287 |
+
},
|
| 288 |
+
"score": 0.9393939393939393,
|
| 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": 0.9414414414414415,
|
| 296 |
+
"score": 0.9414414414414415,
|
| 297 |
+
"score_name": "llama_3_70b_instruct_template_arena_hard"
|
| 298 |
+
},
|
| 299 |
+
"score": 0.9414414414414415,
|
| 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": 0.7999999999999999,
|
| 307 |
+
"f1_Organization": 0.7096774193548386,
|
| 308 |
+
"f1_Location": 0.7058823529411765,
|
| 309 |
+
"f1_macro": 0.7385199240986716,
|
| 310 |
+
"recall_macro": 0.80175983436853,
|
| 311 |
+
"precision_macro": 0.6848220769789397,
|
| 312 |
+
"in_classes_support": 1.0,
|
| 313 |
+
"f1_micro": 0.7361963190184048,
|
| 314 |
+
"recall_micro": 0.8,
|
| 315 |
+
"precision_micro": 0.6818181818181818,
|
| 316 |
+
"score": 0.7361963190184048,
|
| 317 |
+
"score_name": "f1_micro",
|
| 318 |
+
"score_ci_low": 0.6504174815433912,
|
| 319 |
+
"score_ci_high": 0.782581269719831,
|
| 320 |
+
"f1_micro_ci_low": 0.6504174815433912,
|
| 321 |
+
"f1_micro_ci_high": 0.782581269719831
|
| 322 |
+
},
|
| 323 |
+
"score": 0.7361963190184048,
|
| 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": 0.7142857142857143,
|
| 360 |
+
"accuracy_ci_low": 0.2857142857142857,
|
| 361 |
+
"accuracy_ci_high": 1.0,
|
| 362 |
+
"score_name": "accuracy",
|
| 363 |
+
"score": 0.7142857142857143,
|
| 364 |
+
"score_ci_high": 1.0,
|
| 365 |
+
"score_ci_low": 0.2857142857142857,
|
| 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": 0.14285714285714285,
|
| 380 |
+
"accuracy_ci_low": 0.0,
|
| 381 |
+
"accuracy_ci_high": 0.5714285714285714,
|
| 382 |
+
"score_name": "accuracy",
|
| 383 |
+
"score": 0.14285714285714285,
|
| 384 |
+
"score_ci_high": 0.5714285714285714,
|
| 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": 0.8571428571428571,
|
| 410 |
+
"accuracy_ci_low": 0.42857142857142855,
|
| 411 |
+
"accuracy_ci_high": 1.0,
|
| 412 |
+
"score_name": "accuracy",
|
| 413 |
+
"score": 0.8571428571428571,
|
| 414 |
+
"score_ci_high": 1.0,
|
| 415 |
+
"score_ci_low": 0.42857142857142855,
|
| 416 |
+
"num_of_instances": 7
|
| 417 |
+
},
|
| 418 |
+
"mmlu_pro_math": {
|
| 419 |
+
"accuracy": 0.2857142857142857,
|
| 420 |
+
"accuracy_ci_low": 0.0,
|
| 421 |
+
"accuracy_ci_high": 0.7142857142857143,
|
| 422 |
+
"score_name": "accuracy",
|
| 423 |
+
"score": 0.2857142857142857,
|
| 424 |
+
"score_ci_high": 0.7142857142857143,
|
| 425 |
+
"score_ci_low": 0.0,
|
| 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": 0.5714285714285714,
|
| 440 |
+
"accuracy_ci_low": 0.14285714285714285,
|
| 441 |
+
"accuracy_ci_high": 0.8571428571428571,
|
| 442 |
+
"score_name": "accuracy",
|
| 443 |
+
"score": 0.5714285714285714,
|
| 444 |
+
"score_ci_high": 0.8571428571428571,
|
| 445 |
+
"score_ci_low": 0.14285714285714285,
|
| 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": 0.7142857142857143,
|
| 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": 0.45918367346938777,
|
| 469 |
+
"score_name": "subsets_mean",
|
| 470 |
+
"num_of_instances": 98
|
| 471 |
+
},
|
| 472 |
+
"legal": {
|
| 473 |
+
"legalbench_abercrombie": {
|
| 474 |
+
"f1_macro": 0.5199999999999999,
|
| 475 |
+
"f1_suggestive": 0.2,
|
| 476 |
+
"f1_generic": 0.8,
|
| 477 |
+
"f1_fanciful": 0.4,
|
| 478 |
+
"f1_descriptive": 0.4,
|
| 479 |
+
"f1_arbitrary": 0.8,
|
| 480 |
+
"f1_macro_ci_low": 0.3363636363636363,
|
| 481 |
+
"f1_macro_ci_high": 0.766060606060606,
|
| 482 |
+
"score_name": "f1_micro",
|
| 483 |
+
"score": 0.4666666666666667,
|
| 484 |
+
"score_ci_high": 0.6857142857142857,
|
| 485 |
+
"score_ci_low": 0.23076923076923078,
|
| 486 |
+
"num_of_instances": 20,
|
| 487 |
+
"accuracy": 0.35,
|
| 488 |
+
"accuracy_ci_low": 0.15,
|
| 489 |
+
"accuracy_ci_high": 0.6,
|
| 490 |
+
"f1_micro": 0.4666666666666667,
|
| 491 |
+
"f1_micro_ci_low": 0.23076923076923078,
|
| 492 |
+
"f1_micro_ci_high": 0.6857142857142857
|
| 493 |
+
},
|
| 494 |
+
"legalbench_corporate_lobbying": {
|
| 495 |
+
"f1_macro": 0.5793103448275863,
|
| 496 |
+
"f1_no": 0.7586206896551724,
|
| 497 |
+
"f1_yes": 0.4,
|
| 498 |
+
"f1_macro_ci_low": 0.375,
|
| 499 |
+
"f1_macro_ci_high": 0.868365507202327,
|
| 500 |
+
"score_name": "f1_micro",
|
| 501 |
+
"score": 0.6666666666666666,
|
| 502 |
+
"score_ci_high": 0.8717948717948718,
|
| 503 |
+
"score_ci_low": 0.45355819395422325,
|
| 504 |
+
"num_of_instances": 20,
|
| 505 |
+
"accuracy": 0.65,
|
| 506 |
+
"accuracy_ci_low": 0.45,
|
| 507 |
+
"accuracy_ci_high": 0.85,
|
| 508 |
+
"f1_micro": 0.6666666666666666,
|
| 509 |
+
"f1_micro_ci_low": 0.45355819395422325,
|
| 510 |
+
"f1_micro_ci_high": 0.8717948717948718
|
| 511 |
+
},
|
| 512 |
+
"legalbench_function_of_decision_section": {
|
| 513 |
+
"f1_macro": 0.26643990929705214,
|
| 514 |
+
"f1_conclusion": 0.2222222222222222,
|
| 515 |
+
"f1_decree": 0.0,
|
| 516 |
+
"f1_issue": 0.2857142857142857,
|
| 517 |
+
"f1_analysis": 0.5,
|
| 518 |
+
"f1_facts": 0.8571428571428571,
|
| 519 |
+
"f1_procedural history": 0.0,
|
| 520 |
+
"f1_rule": 0.0,
|
| 521 |
+
"f1_macro_ci_low": 0.1142857142857143,
|
| 522 |
+
"f1_macro_ci_high": 0.45052752383482225,
|
| 523 |
+
"score_name": "f1_micro",
|
| 524 |
+
"score": 0.35294117647058826,
|
| 525 |
+
"score_ci_high": 0.5714285714285714,
|
| 526 |
+
"score_ci_low": 0.12903225806451613,
|
| 527 |
+
"num_of_instances": 20,
|
| 528 |
+
"accuracy": 0.3,
|
| 529 |
+
"accuracy_ci_low": 0.1,
|
| 530 |
+
"accuracy_ci_high": 0.5,
|
| 531 |
+
"f1_micro": 0.35294117647058826,
|
| 532 |
+
"f1_micro_ci_low": 0.12903225806451613,
|
| 533 |
+
"f1_micro_ci_high": 0.5714285714285714
|
| 534 |
+
},
|
| 535 |
+
"legalbench_international_citizenship_questions": {
|
| 536 |
+
"f1_macro": 0.696969696969697,
|
| 537 |
+
"f1_yes": 0.7272727272727273,
|
| 538 |
+
"f1_no": 0.6666666666666666,
|
| 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": 0.9,
|
| 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": 0.9,
|
| 549 |
+
"f1_micro": 0.7,
|
| 550 |
+
"f1_micro_ci_low": 0.5,
|
| 551 |
+
"f1_micro_ci_high": 0.9
|
| 552 |
+
},
|
| 553 |
+
"legalbench_proa": {
|
| 554 |
+
"f1_macro": 0.8585526315789473,
|
| 555 |
+
"f1_yes": 0.875,
|
| 556 |
+
"f1_no": 0.8421052631578947,
|
| 557 |
+
"f1_macro_ci_low": 0.6847205623637095,
|
| 558 |
+
"f1_macro_ci_high": 0.9545454545454546,
|
| 559 |
+
"score_name": "f1_micro",
|
| 560 |
+
"score": 0.8571428571428571,
|
| 561 |
+
"score_ci_high": 0.9473684210526315,
|
| 562 |
+
"score_ci_low": 0.6706944990883059,
|
| 563 |
+
"num_of_instances": 20,
|
| 564 |
+
"accuracy": 0.75,
|
| 565 |
+
"accuracy_ci_low": 0.50468235519016,
|
| 566 |
+
"accuracy_ci_high": 0.9,
|
| 567 |
+
"f1_micro": 0.8571428571428571,
|
| 568 |
+
"f1_micro_ci_low": 0.6706944990883059,
|
| 569 |
+
"f1_micro_ci_high": 0.9473684210526315
|
| 570 |
+
},
|
| 571 |
+
"score": 0.6086834733893557,
|
| 572 |
+
"score_name": "subsets_mean",
|
| 573 |
+
"num_of_instances": 100
|
| 574 |
+
},
|
| 575 |
+
"news_classification": {
|
| 576 |
+
"20_newsgroups_short": {
|
| 577 |
+
"f1_macro": 0.6508639236580414,
|
| 578 |
+
"f1_cars": 0.8333333333333334,
|
| 579 |
+
"f1_windows x": 0.3333333333333333,
|
| 580 |
+
"f1_computer graphics": 0.5555555555555556,
|
| 581 |
+
"f1_atheism": 0.2857142857142857,
|
| 582 |
+
"f1_religion": 0.0,
|
| 583 |
+
"f1_medicine": 0.8571428571428571,
|
| 584 |
+
"f1_christianity": 0.8571428571428571,
|
| 585 |
+
"f1_microsoft windows": 0.8333333333333334,
|
| 586 |
+
"f1_middle east": 0.8333333333333334,
|
| 587 |
+
"f1_motorcycles": 0.7272727272727273,
|
| 588 |
+
"f1_pc hardware": 0.7692307692307693,
|
| 589 |
+
"f1_mac hardware": 1.0,
|
| 590 |
+
"f1_electronics": 0.3333333333333333,
|
| 591 |
+
"f1_for sale": 0.75,
|
| 592 |
+
"f1_guns": 0.6666666666666666,
|
| 593 |
+
"f1_space": 0.5714285714285714,
|
| 594 |
+
"f1_cryptography": 0.3333333333333333,
|
| 595 |
+
"f1_baseball": 1.0,
|
| 596 |
+
"f1_politics": 0.5882352941176471,
|
| 597 |
+
"f1_hockey": 0.8888888888888888,
|
| 598 |
+
"f1_macro_ci_low": 0.5563237523685153,
|
| 599 |
+
"f1_macro_ci_high": 0.7401830936884861,
|
| 600 |
+
"score_name": "f1_micro",
|
| 601 |
+
"score": 0.6907216494845361,
|
| 602 |
+
"score_ci_high": 0.7645361476644139,
|
| 603 |
+
"score_ci_low": 0.5804059043570197,
|
| 604 |
+
"num_of_instances": 100,
|
| 605 |
+
"accuracy": 0.67,
|
| 606 |
+
"accuracy_ci_low": 0.56,
|
| 607 |
+
"accuracy_ci_high": 0.75,
|
| 608 |
+
"f1_micro": 0.6907216494845361,
|
| 609 |
+
"f1_micro_ci_low": 0.5804059043570197,
|
| 610 |
+
"f1_micro_ci_high": 0.7645361476644139
|
| 611 |
+
},
|
| 612 |
+
"score": 0.6907216494845361,
|
| 613 |
+
"score_name": "subsets_mean",
|
| 614 |
+
"num_of_instances": 100
|
| 615 |
+
},
|
| 616 |
+
"product_help": {
|
| 617 |
+
"cfpb_product_2023": {
|
| 618 |
+
"f1_macro": 0.7410023219814241,
|
| 619 |
+
"f1_credit reporting or credit repair services or other personal consumer reports": 0.9411764705882353,
|
| 620 |
+
"f1_debt collection": 0.7368421052631579,
|
| 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": 1.0,
|
| 625 |
+
"f1_mortgage": 0.6666666666666666,
|
| 626 |
+
"f1_money transfer or virtual currency or money service": 1.0,
|
| 627 |
+
"f1_macro_ci_low": 0.548026288422304,
|
| 628 |
+
"f1_macro_ci_high": 0.8446388968731338,
|
| 629 |
+
"score_name": "f1_micro",
|
| 630 |
+
"score": 0.898989898989899,
|
| 631 |
+
"score_ci_high": 0.9447236180904522,
|
| 632 |
+
"score_ci_low": 0.8203180754561684,
|
| 633 |
+
"num_of_instances": 100,
|
| 634 |
+
"accuracy": 0.89,
|
| 635 |
+
"accuracy_ci_low": 0.81,
|
| 636 |
+
"accuracy_ci_high": 0.94,
|
| 637 |
+
"f1_micro": 0.898989898989899,
|
| 638 |
+
"f1_micro_ci_low": 0.8203180754561684,
|
| 639 |
+
"f1_micro_ci_high": 0.9447236180904522
|
| 640 |
+
},
|
| 641 |
+
"cfpb_product_watsonx": {
|
| 642 |
+
"f1_macro": 0.8260549519130755,
|
| 643 |
+
"f1_mortgages and loans": 0.8695652173913043,
|
| 644 |
+
"f1_credit card": 0.8181818181818182,
|
| 645 |
+
"f1_debt collection": 0.7368421052631579,
|
| 646 |
+
"f1_credit reporting": 0.782608695652174,
|
| 647 |
+
"f1_retail banking": 0.9230769230769231,
|
| 648 |
+
"f1_macro_ci_low": 0.7025385602172309,
|
| 649 |
+
"f1_macro_ci_high": 0.9116966675085018,
|
| 650 |
+
"score_name": "f1_micro",
|
| 651 |
+
"score": 0.82,
|
| 652 |
+
"score_ci_high": 0.9,
|
| 653 |
+
"score_ci_low": 0.7,
|
| 654 |
+
"num_of_instances": 50,
|
| 655 |
+
"accuracy": 0.82,
|
| 656 |
+
"accuracy_ci_low": 0.7,
|
| 657 |
+
"accuracy_ci_high": 0.9,
|
| 658 |
+
"f1_micro": 0.82,
|
| 659 |
+
"f1_micro_ci_low": 0.7,
|
| 660 |
+
"f1_micro_ci_high": 0.9
|
| 661 |
+
},
|
| 662 |
+
"score": 0.8594949494949495,
|
| 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": 0.36,
|
| 670 |
+
"program_accuracy": 0.37,
|
| 671 |
+
"score": 0.37,
|
| 672 |
+
"score_name": "program_accuracy",
|
| 673 |
+
"execution_accuracy_ci_low": 0.27,
|
| 674 |
+
"execution_accuracy_ci_high": 0.47,
|
| 675 |
+
"program_accuracy_ci_low": 0.2811354521803329,
|
| 676 |
+
"program_accuracy_ci_high": 0.47,
|
| 677 |
+
"score_ci_low": 0.2811354521803329,
|
| 678 |
+
"score_ci_high": 0.47
|
| 679 |
+
},
|
| 680 |
+
"score": 0.37,
|
| 681 |
+
"score_name": "subsets_mean",
|
| 682 |
+
"num_of_instances": 100
|
| 683 |
+
},
|
| 684 |
+
"rag_general": {
|
| 685 |
+
"rag_response_generation_clapnq": {
|
| 686 |
+
"precision": 0.4468480048160072,
|
| 687 |
+
"recall": 0.6454220314570971,
|
| 688 |
+
"f1": 0.4890230179217559,
|
| 689 |
+
"precision_ci_low": 0.4151579792947102,
|
| 690 |
+
"precision_ci_high": 0.4845603023826521,
|
| 691 |
+
"recall_ci_low": 0.6061330597384731,
|
| 692 |
+
"recall_ci_high": 0.6765859073642304,
|
| 693 |
+
"f1_ci_low": 0.46005008735451697,
|
| 694 |
+
"f1_ci_high": 0.5174377148355835,
|
| 695 |
+
"score_name": "f1",
|
| 696 |
+
"score": 0.4890230179217559,
|
| 697 |
+
"score_ci_high": 0.5174377148355835,
|
| 698 |
+
"score_ci_low": 0.46005008735451697,
|
| 699 |
+
"num_of_instances": 100,
|
| 700 |
+
"correctness_f1_bert_score.deberta_large_mnli": 0.691885244846344,
|
| 701 |
+
"correctness_recall_bert_score.deberta_large_mnli": 0.7364352762699127,
|
| 702 |
+
"correctness_precision_bert_score.deberta_large_mnli": 0.6613857835531235,
|
| 703 |
+
"faithfullness_f1_token_overlap": 0.43039423317750836,
|
| 704 |
+
"faithfullness_recall_token_overlap": 0.35370116631248527,
|
| 705 |
+
"faithfullness_precision_token_overlap": 0.6790419566943843,
|
| 706 |
+
"correctness_f1_token_overlap": 0.4890230179217559,
|
| 707 |
+
"correctness_recall_token_overlap": 0.6454220314570971,
|
| 708 |
+
"correctness_precision_token_overlap": 0.4468480048160072
|
| 709 |
+
},
|
| 710 |
+
"score": 0.4890230179217559,
|
| 711 |
+
"score_name": "subsets_mean",
|
| 712 |
+
"num_of_instances": 100
|
| 713 |
+
},
|
| 714 |
+
"reasoning": {
|
| 715 |
+
"hellaswag": {
|
| 716 |
+
"accuracy": 0.62,
|
| 717 |
+
"accuracy_ci_low": 0.51,
|
| 718 |
+
"accuracy_ci_high": 0.7,
|
| 719 |
+
"score_name": "accuracy",
|
| 720 |
+
"score": 0.62,
|
| 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": 0.89,
|
| 728 |
+
"accuracy_ci_high": 0.98,
|
| 729 |
+
"score_name": "accuracy",
|
| 730 |
+
"score": 0.94,
|
| 731 |
+
"score_ci_high": 0.98,
|
| 732 |
+
"score_ci_low": 0.89,
|
| 733 |
+
"num_of_instances": 100
|
| 734 |
+
},
|
| 735 |
+
"score": 0.78,
|
| 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": 0.8862442294401782,
|
| 746 |
+
"category_deception": 0.9089467666215367,
|
| 747 |
+
"category_explicit_content": 0.8660169005243467,
|
| 748 |
+
"category_violence": 0.8813284475474604,
|
| 749 |
+
"category_discrimination": 0.7921433370737802,
|
| 750 |
+
"category_substance_abuse": 0.892341727639238,
|
| 751 |
+
"category_pii": 0.8837605582343208,
|
| 752 |
+
"safety": 0.8784392809950642,
|
| 753 |
+
"safety_ci_low": 0.8622394069744699,
|
| 754 |
+
"safety_ci_high": 0.8922010585173956,
|
| 755 |
+
"score_name": "safety",
|
| 756 |
+
"score": 0.8784392809950642,
|
| 757 |
+
"score_ci_high": 0.8922010585173956,
|
| 758 |
+
"score_ci_low": 0.8622394069744699,
|
| 759 |
+
"num_of_instances": 100
|
| 760 |
+
},
|
| 761 |
+
"score": 0.8784392809950642,
|
| 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": 0.3334800194661833,
|
| 769 |
+
"rouge2": 0.1746055052600959,
|
| 770 |
+
"rouge1": 0.3914063192107438,
|
| 771 |
+
"rougeL": 0.2607874086538398,
|
| 772 |
+
"score": 0.2607874086538398,
|
| 773 |
+
"score_name": "rougeL",
|
| 774 |
+
"rougeLsum_ci_low": 0.31372911043552615,
|
| 775 |
+
"rougeLsum_ci_high": 0.3521588230191023,
|
| 776 |
+
"rouge2_ci_low": 0.16154343447204916,
|
| 777 |
+
"rouge2_ci_high": 0.18686015868835737,
|
| 778 |
+
"rouge1_ci_low": 0.36979472683613324,
|
| 779 |
+
"rouge1_ci_high": 0.41112417139895086,
|
| 780 |
+
"rougeL_ci_low": 0.24640816277583857,
|
| 781 |
+
"rougeL_ci_high": 0.2769615579591105,
|
| 782 |
+
"score_ci_low": 0.24640816277583857,
|
| 783 |
+
"score_ci_high": 0.2769615579591105
|
| 784 |
+
},
|
| 785 |
+
"tldr_document_filtered_to_6000_chars": {
|
| 786 |
+
"num_of_instances": 100,
|
| 787 |
+
"rougeLsum": 0.08980268920854961,
|
| 788 |
+
"rouge2": 0.012765072464340065,
|
| 789 |
+
"rouge1": 0.10813972613064696,
|
| 790 |
+
"rougeL": 0.08042515106945874,
|
| 791 |
+
"score": 0.08042515106945874,
|
| 792 |
+
"score_name": "rougeL",
|
| 793 |
+
"rougeLsum_ci_low": 0.07759889049057624,
|
| 794 |
+
"rougeLsum_ci_high": 0.1032183831189881,
|
| 795 |
+
"rouge2_ci_low": 0.008976164023617238,
|
| 796 |
+
"rouge2_ci_high": 0.018103937102878734,
|
| 797 |
+
"rouge1_ci_low": 0.09248155246218394,
|
| 798 |
+
"rouge1_ci_high": 0.1250184504642972,
|
| 799 |
+
"rougeL_ci_low": 0.06959906992080947,
|
| 800 |
+
"rougeL_ci_high": 0.09223760457164616,
|
| 801 |
+
"score_ci_low": 0.06959906992080947,
|
| 802 |
+
"score_ci_high": 0.09223760457164616
|
| 803 |
+
},
|
| 804 |
+
"score": 0.17060627986164928,
|
| 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 |
+
156,
|
| 813 |
+
105,
|
| 814 |
+
71,
|
| 815 |
+
50
|
| 816 |
+
],
|
| 817 |
+
"totals": [
|
| 818 |
+
226,
|
| 819 |
+
220,
|
| 820 |
+
214,
|
| 821 |
+
208
|
| 822 |
+
],
|
| 823 |
+
"precisions": [
|
| 824 |
+
0.6902654867256638,
|
| 825 |
+
0.4772727272727273,
|
| 826 |
+
0.3317757009345794,
|
| 827 |
+
0.2403846153846154
|
| 828 |
+
],
|
| 829 |
+
"bp": 1.0,
|
| 830 |
+
"sys_len": 226,
|
| 831 |
+
"ref_len": 208,
|
| 832 |
+
"sacrebleu": 0.4026090246453075,
|
| 833 |
+
"score": 0.4026090246453075,
|
| 834 |
+
"score_name": "sacrebleu",
|
| 835 |
+
"score_ci_low": 0.24025628806486451,
|
| 836 |
+
"score_ci_high": 0.5204212192812463,
|
| 837 |
+
"sacrebleu_ci_low": 0.24025628806486451,
|
| 838 |
+
"sacrebleu_ci_high": 0.5204212192812463
|
| 839 |
+
},
|
| 840 |
+
"mt_flores_101_deu_eng": {
|
| 841 |
+
"num_of_instances": 6,
|
| 842 |
+
"counts": [
|
| 843 |
+
143,
|
| 844 |
+
93,
|
| 845 |
+
63,
|
| 846 |
+
45
|
| 847 |
+
],
|
| 848 |
+
"totals": [
|
| 849 |
+
219,
|
| 850 |
+
213,
|
| 851 |
+
207,
|
| 852 |
+
201
|
| 853 |
+
],
|
| 854 |
+
"precisions": [
|
| 855 |
+
0.6529680365296804,
|
| 856 |
+
0.43661971830985913,
|
| 857 |
+
0.30434782608695654,
|
| 858 |
+
0.22388059701492538
|
| 859 |
+
],
|
| 860 |
+
"bp": 1.0,
|
| 861 |
+
"sys_len": 219,
|
| 862 |
+
"ref_len": 208,
|
| 863 |
+
"sacrebleu": 0.3733322279107499,
|
| 864 |
+
"score": 0.3733322279107499,
|
| 865 |
+
"score_name": "sacrebleu",
|
| 866 |
+
"score_ci_low": 0.26773105555308707,
|
| 867 |
+
"score_ci_high": 0.5628718725153604,
|
| 868 |
+
"sacrebleu_ci_low": 0.26773105555308707,
|
| 869 |
+
"sacrebleu_ci_high": 0.5628718725153604
|
| 870 |
+
},
|
| 871 |
+
"mt_flores_101_eng_ara": {
|
| 872 |
+
"num_of_instances": 6,
|
| 873 |
+
"counts": [
|
| 874 |
+
118,
|
| 875 |
+
70,
|
| 876 |
+
45,
|
| 877 |
+
27
|
| 878 |
+
],
|
| 879 |
+
"totals": [
|
| 880 |
+
211,
|
| 881 |
+
205,
|
| 882 |
+
199,
|
| 883 |
+
193
|
| 884 |
+
],
|
| 885 |
+
"precisions": [
|
| 886 |
+
0.5592417061611374,
|
| 887 |
+
0.34146341463414637,
|
| 888 |
+
0.22613065326633167,
|
| 889 |
+
0.13989637305699482
|
| 890 |
+
],
|
| 891 |
+
"bp": 1.0,
|
| 892 |
+
"sys_len": 211,
|
| 893 |
+
"ref_len": 209,
|
| 894 |
+
"sacrebleu": 0.27879013737554453,
|
| 895 |
+
"score": 0.27879013737554453,
|
| 896 |
+
"score_name": "sacrebleu",
|
| 897 |
+
"score_ci_low": 0.242285190601197,
|
| 898 |
+
"score_ci_high": 0.3370610364230129,
|
| 899 |
+
"sacrebleu_ci_low": 0.242285190601197,
|
| 900 |
+
"sacrebleu_ci_high": 0.3370610364230129
|
| 901 |
+
},
|
| 902 |
+
"mt_flores_101_eng_deu": {
|
| 903 |
+
"num_of_instances": 6,
|
| 904 |
+
"counts": [
|
| 905 |
+
148,
|
| 906 |
+
92,
|
| 907 |
+
62,
|
| 908 |
+
46
|
| 909 |
+
],
|
| 910 |
+
"totals": [
|
| 911 |
+
217,
|
| 912 |
+
211,
|
| 913 |
+
205,
|
| 914 |
+
199
|
| 915 |
+
],
|
| 916 |
+
"precisions": [
|
| 917 |
+
0.6820276497695852,
|
| 918 |
+
0.43601895734597157,
|
| 919 |
+
0.3024390243902439,
|
| 920 |
+
0.23115577889447236
|
| 921 |
+
],
|
| 922 |
+
"bp": 1.0,
|
| 923 |
+
"sys_len": 217,
|
| 924 |
+
"ref_len": 216,
|
| 925 |
+
"sacrebleu": 0.3797191362653853,
|
| 926 |
+
"score": 0.3797191362653853,
|
| 927 |
+
"score_name": "sacrebleu",
|
| 928 |
+
"score_ci_low": 0.24991245830969483,
|
| 929 |
+
"score_ci_high": 0.537890541133251,
|
| 930 |
+
"sacrebleu_ci_low": 0.24991245830969483,
|
| 931 |
+
"sacrebleu_ci_high": 0.537890541133251
|
| 932 |
+
},
|
| 933 |
+
"mt_flores_101_eng_fra": {
|
| 934 |
+
"num_of_instances": 6,
|
| 935 |
+
"counts": [
|
| 936 |
+
190,
|
| 937 |
+
150,
|
| 938 |
+
119,
|
| 939 |
+
95
|
| 940 |
+
],
|
| 941 |
+
"totals": [
|
| 942 |
+
244,
|
| 943 |
+
238,
|
| 944 |
+
232,
|
| 945 |
+
226
|
| 946 |
+
],
|
| 947 |
+
"precisions": [
|
| 948 |
+
0.7786885245901639,
|
| 949 |
+
0.6302521008403361,
|
| 950 |
+
0.5129310344827587,
|
| 951 |
+
0.42035398230088494
|
| 952 |
+
],
|
| 953 |
+
"bp": 1.0,
|
| 954 |
+
"sys_len": 244,
|
| 955 |
+
"ref_len": 235,
|
| 956 |
+
"sacrebleu": 0.5703455465960385,
|
| 957 |
+
"score": 0.5703455465960385,
|
| 958 |
+
"score_name": "sacrebleu",
|
| 959 |
+
"score_ci_low": 0.4931990672262396,
|
| 960 |
+
"score_ci_high": 0.6753511173840523,
|
| 961 |
+
"sacrebleu_ci_low": 0.4931990672262396,
|
| 962 |
+
"sacrebleu_ci_high": 0.6753511173840523
|
| 963 |
+
},
|
| 964 |
+
"mt_flores_101_eng_kor": {
|
| 965 |
+
"num_of_instances": 6,
|
| 966 |
+
"counts": [
|
| 967 |
+
162,
|
| 968 |
+
92,
|
| 969 |
+
63,
|
| 970 |
+
43
|
| 971 |
+
],
|
| 972 |
+
"totals": [
|
| 973 |
+
273,
|
| 974 |
+
267,
|
| 975 |
+
261,
|
| 976 |
+
255
|
| 977 |
+
],
|
| 978 |
+
"precisions": [
|
| 979 |
+
0.5934065934065934,
|
| 980 |
+
0.3445692883895131,
|
| 981 |
+
0.24137931034482757,
|
| 982 |
+
0.16862745098039217
|
| 983 |
+
],
|
| 984 |
+
"bp": 1.0,
|
| 985 |
+
"sys_len": 273,
|
| 986 |
+
"ref_len": 249,
|
| 987 |
+
"sacrebleu": 0.3020398964371346,
|
| 988 |
+
"score": 0.3020398964371346,
|
| 989 |
+
"score_name": "sacrebleu",
|
| 990 |
+
"score_ci_low": 0.23699249066795575,
|
| 991 |
+
"score_ci_high": 0.36104670705975184,
|
| 992 |
+
"sacrebleu_ci_low": 0.23699249066795575,
|
| 993 |
+
"sacrebleu_ci_high": 0.36104670705975184
|
| 994 |
+
},
|
| 995 |
+
"mt_flores_101_eng_por": {
|
| 996 |
+
"num_of_instances": 6,
|
| 997 |
+
"counts": [
|
| 998 |
+
181,
|
| 999 |
+
136,
|
| 1000 |
+
110,
|
| 1001 |
+
89
|
| 1002 |
+
],
|
| 1003 |
+
"totals": [
|
| 1004 |
+
228,
|
| 1005 |
+
222,
|
| 1006 |
+
216,
|
| 1007 |
+
210
|
| 1008 |
+
],
|
| 1009 |
+
"precisions": [
|
| 1010 |
+
0.7938596491228069,
|
| 1011 |
+
0.6126126126126126,
|
| 1012 |
+
0.5092592592592592,
|
| 1013 |
+
0.4238095238095238
|
| 1014 |
+
],
|
| 1015 |
+
"bp": 1.0,
|
| 1016 |
+
"sys_len": 228,
|
| 1017 |
+
"ref_len": 222,
|
| 1018 |
+
"sacrebleu": 0.5691933623646763,
|
| 1019 |
+
"score": 0.5691933623646763,
|
| 1020 |
+
"score_name": "sacrebleu",
|
| 1021 |
+
"score_ci_low": 0.5074329158262365,
|
| 1022 |
+
"score_ci_high": 0.6529930317459868,
|
| 1023 |
+
"sacrebleu_ci_low": 0.5074329158262365,
|
| 1024 |
+
"sacrebleu_ci_high": 0.6529930317459868
|
| 1025 |
+
},
|
| 1026 |
+
"mt_flores_101_eng_ron": {
|
| 1027 |
+
"num_of_instances": 6,
|
| 1028 |
+
"counts": [
|
| 1029 |
+
165,
|
| 1030 |
+
125,
|
| 1031 |
+
97,
|
| 1032 |
+
82
|
| 1033 |
+
],
|
| 1034 |
+
"totals": [
|
| 1035 |
+
230,
|
| 1036 |
+
224,
|
| 1037 |
+
218,
|
| 1038 |
+
212
|
| 1039 |
+
],
|
| 1040 |
+
"precisions": [
|
| 1041 |
+
0.717391304347826,
|
| 1042 |
+
0.5580357142857143,
|
| 1043 |
+
0.444954128440367,
|
| 1044 |
+
0.38679245283018865
|
| 1045 |
+
],
|
| 1046 |
+
"bp": 1.0,
|
| 1047 |
+
"sys_len": 230,
|
| 1048 |
+
"ref_len": 230,
|
| 1049 |
+
"sacrebleu": 0.5123335940057543,
|
| 1050 |
+
"score": 0.5123335940057543,
|
| 1051 |
+
"score_name": "sacrebleu",
|
| 1052 |
+
"score_ci_low": 0.3840244184926614,
|
| 1053 |
+
"score_ci_high": 0.650779271851941,
|
| 1054 |
+
"sacrebleu_ci_low": 0.3840244184926614,
|
| 1055 |
+
"sacrebleu_ci_high": 0.650779271851941
|
| 1056 |
+
},
|
| 1057 |
+
"mt_flores_101_eng_spa": {
|
| 1058 |
+
"num_of_instances": 6,
|
| 1059 |
+
"counts": [
|
| 1060 |
+
170,
|
| 1061 |
+
109,
|
| 1062 |
+
73,
|
| 1063 |
+
47
|
| 1064 |
+
],
|
| 1065 |
+
"totals": [
|
| 1066 |
+
236,
|
| 1067 |
+
230,
|
| 1068 |
+
224,
|
| 1069 |
+
218
|
| 1070 |
+
],
|
| 1071 |
+
"precisions": [
|
| 1072 |
+
0.7203389830508474,
|
| 1073 |
+
0.47391304347826085,
|
| 1074 |
+
0.32589285714285715,
|
| 1075 |
+
0.21559633027522934
|
| 1076 |
+
],
|
| 1077 |
+
"bp": 0.9707745538991623,
|
| 1078 |
+
"sys_len": 236,
|
| 1079 |
+
"ref_len": 243,
|
| 1080 |
+
"sacrebleu": 0.3820377957320921,
|
| 1081 |
+
"score": 0.3820377957320921,
|
| 1082 |
+
"score_name": "sacrebleu",
|
| 1083 |
+
"score_ci_low": 0.3085199315474183,
|
| 1084 |
+
"score_ci_high": 0.48773289703817563,
|
| 1085 |
+
"sacrebleu_ci_low": 0.3085199315474183,
|
| 1086 |
+
"sacrebleu_ci_high": 0.48773289703817563
|
| 1087 |
+
},
|
| 1088 |
+
"mt_flores_101_fra_eng": {
|
| 1089 |
+
"num_of_instances": 6,
|
| 1090 |
+
"counts": [
|
| 1091 |
+
158,
|
| 1092 |
+
111,
|
| 1093 |
+
79,
|
| 1094 |
+
59
|
| 1095 |
+
],
|
| 1096 |
+
"totals": [
|
| 1097 |
+
225,
|
| 1098 |
+
219,
|
| 1099 |
+
213,
|
| 1100 |
+
207
|
| 1101 |
+
],
|
| 1102 |
+
"precisions": [
|
| 1103 |
+
0.7022222222222223,
|
| 1104 |
+
0.5068493150684932,
|
| 1105 |
+
0.37089201877934275,
|
| 1106 |
+
0.28502415458937197
|
| 1107 |
+
],
|
| 1108 |
+
"bp": 1.0,
|
| 1109 |
+
"sys_len": 225,
|
| 1110 |
+
"ref_len": 208,
|
| 1111 |
+
"sacrebleu": 0.44042366511701625,
|
| 1112 |
+
"score": 0.44042366511701625,
|
| 1113 |
+
"score_name": "sacrebleu",
|
| 1114 |
+
"score_ci_low": 0.341980662697259,
|
| 1115 |
+
"score_ci_high": 0.5496900259922087,
|
| 1116 |
+
"sacrebleu_ci_low": 0.341980662697259,
|
| 1117 |
+
"sacrebleu_ci_high": 0.5496900259922087
|
| 1118 |
+
},
|
| 1119 |
+
"mt_flores_101_jpn_eng": {
|
| 1120 |
+
"num_of_instances": 6,
|
| 1121 |
+
"counts": [
|
| 1122 |
+
150,
|
| 1123 |
+
90,
|
| 1124 |
+
58,
|
| 1125 |
+
39
|
| 1126 |
+
],
|
| 1127 |
+
"totals": [
|
| 1128 |
+
236,
|
| 1129 |
+
230,
|
| 1130 |
+
224,
|
| 1131 |
+
218
|
| 1132 |
+
],
|
| 1133 |
+
"precisions": [
|
| 1134 |
+
0.635593220338983,
|
| 1135 |
+
0.391304347826087,
|
| 1136 |
+
0.2589285714285714,
|
| 1137 |
+
0.17889908256880735
|
| 1138 |
+
],
|
| 1139 |
+
"bp": 1.0,
|
| 1140 |
+
"sys_len": 236,
|
| 1141 |
+
"ref_len": 208,
|
| 1142 |
+
"sacrebleu": 0.32762007434781915,
|
| 1143 |
+
"score": 0.32762007434781915,
|
| 1144 |
+
"score_name": "sacrebleu",
|
| 1145 |
+
"score_ci_low": 0.20337120135607298,
|
| 1146 |
+
"score_ci_high": 0.4231259198402102,
|
| 1147 |
+
"sacrebleu_ci_low": 0.20337120135607298,
|
| 1148 |
+
"sacrebleu_ci_high": 0.4231259198402102
|
| 1149 |
+
},
|
| 1150 |
+
"mt_flores_101_kor_eng": {
|
| 1151 |
+
"num_of_instances": 6,
|
| 1152 |
+
"counts": [
|
| 1153 |
+
135,
|
| 1154 |
+
76,
|
| 1155 |
+
45,
|
| 1156 |
+
29
|
| 1157 |
+
],
|
| 1158 |
+
"totals": [
|
| 1159 |
+
226,
|
| 1160 |
+
220,
|
| 1161 |
+
214,
|
| 1162 |
+
208
|
| 1163 |
+
],
|
| 1164 |
+
"precisions": [
|
| 1165 |
+
0.5973451327433628,
|
| 1166 |
+
0.34545454545454546,
|
| 1167 |
+
0.2102803738317757,
|
| 1168 |
+
0.13942307692307693
|
| 1169 |
+
],
|
| 1170 |
+
"bp": 1.0,
|
| 1171 |
+
"sys_len": 226,
|
| 1172 |
+
"ref_len": 208,
|
| 1173 |
+
"sacrebleu": 0.2788928697008729,
|
| 1174 |
+
"score": 0.2788928697008729,
|
| 1175 |
+
"score_name": "sacrebleu",
|
| 1176 |
+
"score_ci_low": 0.18358949678749556,
|
| 1177 |
+
"score_ci_high": 0.4387198667737409,
|
| 1178 |
+
"sacrebleu_ci_low": 0.18358949678749556,
|
| 1179 |
+
"sacrebleu_ci_high": 0.4387198667737409
|
| 1180 |
+
},
|
| 1181 |
+
"mt_flores_101_por_eng": {
|
| 1182 |
+
"num_of_instances": 6,
|
| 1183 |
+
"counts": [
|
| 1184 |
+
168,
|
| 1185 |
+
124,
|
| 1186 |
+
93,
|
| 1187 |
+
73
|
| 1188 |
+
],
|
| 1189 |
+
"totals": [
|
| 1190 |
+
221,
|
| 1191 |
+
215,
|
| 1192 |
+
209,
|
| 1193 |
+
203
|
| 1194 |
+
],
|
| 1195 |
+
"precisions": [
|
| 1196 |
+
0.7601809954751131,
|
| 1197 |
+
0.5767441860465117,
|
| 1198 |
+
0.4449760765550239,
|
| 1199 |
+
0.35960591133004927
|
| 1200 |
+
],
|
| 1201 |
+
"bp": 1.0,
|
| 1202 |
+
"sys_len": 221,
|
| 1203 |
+
"ref_len": 208,
|
| 1204 |
+
"sacrebleu": 0.5146546836832124,
|
| 1205 |
+
"score": 0.5146546836832124,
|
| 1206 |
+
"score_name": "sacrebleu",
|
| 1207 |
+
"score_ci_low": 0.378423475348725,
|
| 1208 |
+
"score_ci_high": 0.6374865872050556,
|
| 1209 |
+
"sacrebleu_ci_low": 0.378423475348725,
|
| 1210 |
+
"sacrebleu_ci_high": 0.6374865872050556
|
| 1211 |
+
},
|
| 1212 |
+
"mt_flores_101_ron_eng": {
|
| 1213 |
+
"num_of_instances": 6,
|
| 1214 |
+
"counts": [
|
| 1215 |
+
166,
|
| 1216 |
+
121,
|
| 1217 |
+
91,
|
| 1218 |
+
72
|
| 1219 |
+
],
|
| 1220 |
+
"totals": [
|
| 1221 |
+
227,
|
| 1222 |
+
221,
|
| 1223 |
+
215,
|
| 1224 |
+
209
|
| 1225 |
+
],
|
| 1226 |
+
"precisions": [
|
| 1227 |
+
0.7312775330396475,
|
| 1228 |
+
0.5475113122171945,
|
| 1229 |
+
0.4232558139534883,
|
| 1230 |
+
0.34449760765550236
|
| 1231 |
+
],
|
| 1232 |
+
"bp": 1.0,
|
| 1233 |
+
"sys_len": 227,
|
| 1234 |
+
"ref_len": 208,
|
| 1235 |
+
"sacrebleu": 0.49154820843517344,
|
| 1236 |
+
"score": 0.49154820843517344,
|
| 1237 |
+
"score_name": "sacrebleu",
|
| 1238 |
+
"score_ci_low": 0.4117068924616905,
|
| 1239 |
+
"score_ci_high": 0.6461238952977133,
|
| 1240 |
+
"sacrebleu_ci_low": 0.4117068924616905,
|
| 1241 |
+
"sacrebleu_ci_high": 0.6461238952977133
|
| 1242 |
+
},
|
| 1243 |
+
"mt_flores_101_spa_eng": {
|
| 1244 |
+
"num_of_instances": 6,
|
| 1245 |
+
"counts": [
|
| 1246 |
+
150,
|
| 1247 |
+
96,
|
| 1248 |
+
65,
|
| 1249 |
+
48
|
| 1250 |
+
],
|
| 1251 |
+
"totals": [
|
| 1252 |
+
219,
|
| 1253 |
+
213,
|
| 1254 |
+
207,
|
| 1255 |
+
201
|
| 1256 |
+
],
|
| 1257 |
+
"precisions": [
|
| 1258 |
+
0.684931506849315,
|
| 1259 |
+
0.45070422535211263,
|
| 1260 |
+
0.3140096618357488,
|
| 1261 |
+
0.23880597014925375
|
| 1262 |
+
],
|
| 1263 |
+
"bp": 1.0,
|
| 1264 |
+
"sys_len": 219,
|
| 1265 |
+
"ref_len": 208,
|
| 1266 |
+
"sacrebleu": 0.3900602917326207,
|
| 1267 |
+
"score": 0.3900602917326207,
|
| 1268 |
+
"score_name": "sacrebleu",
|
| 1269 |
+
"score_ci_low": 0.3529954080391477,
|
| 1270 |
+
"score_ci_high": 0.44601468942834244,
|
| 1271 |
+
"sacrebleu_ci_low": 0.3529954080391477,
|
| 1272 |
+
"sacrebleu_ci_high": 0.44601468942834244
|
| 1273 |
+
},
|
| 1274 |
+
"score": 0.41424003428995987,
|
| 1275 |
+
"score_name": "subsets_mean",
|
| 1276 |
+
"num_of_instances": 90
|
| 1277 |
+
},
|
| 1278 |
+
"score": 0.6413403122123419,
|
| 1279 |
+
"score_name": "subsets_mean",
|
| 1280 |
+
"num_of_instances": 1537
|
| 1281 |
+
}
|
| 1282 |
+
}
|
results/bluebench/2025-08-03T12-21-28_evaluation_results.json
ADDED
|
@@ -0,0 +1,1282 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"environment_info": {
|
| 3 |
+
"timestamp_utc": "2025-08-03T16:21:24.530955Z",
|
| 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/o3-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/o3-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": 0.8888888888888888,
|
| 180 |
+
"accuracy_ci_low": 0.5310928992288233,
|
| 181 |
+
"accuracy_ci_high": 1.0,
|
| 182 |
+
"score_name": "accuracy",
|
| 183 |
+
"score": 0.8888888888888888,
|
| 184 |
+
"score_ci_high": 1.0,
|
| 185 |
+
"score_ci_low": 0.5310928992288233,
|
| 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": 0.8888888888888888,
|
| 200 |
+
"accuracy_ci_low": 0.46041936253217447,
|
| 201 |
+
"accuracy_ci_high": 1.0,
|
| 202 |
+
"score_name": "accuracy",
|
| 203 |
+
"score": 0.8888888888888888,
|
| 204 |
+
"score_ci_high": 1.0,
|
| 205 |
+
"score_ci_low": 0.46041936253217447,
|
| 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": 1.0,
|
| 260 |
+
"accuracy_ci_low": 1.0,
|
| 261 |
+
"accuracy_ci_high": 1.0,
|
| 262 |
+
"score_name": "accuracy",
|
| 263 |
+
"score": 1.0,
|
| 264 |
+
"score_ci_high": 1.0,
|
| 265 |
+
"score_ci_low": 1.0,
|
| 266 |
+
"num_of_instances": 9
|
| 267 |
+
},
|
| 268 |
+
"safety_bbq_ses": {
|
| 269 |
+
"accuracy": 1.0,
|
| 270 |
+
"accuracy_ci_low": 1.0,
|
| 271 |
+
"accuracy_ci_high": 1.0,
|
| 272 |
+
"score_name": "accuracy",
|
| 273 |
+
"score": 1.0,
|
| 274 |
+
"score_ci_high": 1.0,
|
| 275 |
+
"score_ci_low": 1.0,
|
| 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": 0.3711340206185567,
|
| 296 |
+
"score": 0.3711340206185567,
|
| 297 |
+
"score_name": "llama_3_70b_instruct_template_arena_hard"
|
| 298 |
+
},
|
| 299 |
+
"score": 0.3711340206185567,
|
| 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": 0.8181818181818182,
|
| 307 |
+
"f1_Organization": 0.6428571428571429,
|
| 308 |
+
"f1_Location": 0.75,
|
| 309 |
+
"f1_macro": 0.737012987012987,
|
| 310 |
+
"recall_macro": 0.683488612836439,
|
| 311 |
+
"precision_macro": 0.8125,
|
| 312 |
+
"in_classes_support": 1.0,
|
| 313 |
+
"f1_micro": 0.7285714285714285,
|
| 314 |
+
"recall_micro": 0.68,
|
| 315 |
+
"precision_micro": 0.7846153846153846,
|
| 316 |
+
"score": 0.7285714285714285,
|
| 317 |
+
"score_name": "f1_micro",
|
| 318 |
+
"score_ci_low": 0.5888115359844259,
|
| 319 |
+
"score_ci_high": 0.8223830090806191,
|
| 320 |
+
"f1_micro_ci_low": 0.5888115359844259,
|
| 321 |
+
"f1_micro_ci_high": 0.8223830090806191
|
| 322 |
+
},
|
| 323 |
+
"score": 0.7285714285714285,
|
| 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": 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": 0.7142857142857143,
|
| 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": 0.42857142857142855,
|
| 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": 0.2857142857142857,
|
| 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": 0.14285714285714285,
|
| 400 |
+
"accuracy_ci_low": 0.0,
|
| 401 |
+
"accuracy_ci_high": 0.6807203593841678,
|
| 402 |
+
"score_name": "accuracy",
|
| 403 |
+
"score": 0.14285714285714285,
|
| 404 |
+
"score_ci_high": 0.6807203593841678,
|
| 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": 1.0,
|
| 420 |
+
"accuracy_ci_low": 1.0,
|
| 421 |
+
"accuracy_ci_high": 1.0,
|
| 422 |
+
"score_name": "accuracy",
|
| 423 |
+
"score": 1.0,
|
| 424 |
+
"score_ci_high": 1.0,
|
| 425 |
+
"score_ci_low": 1.0,
|
| 426 |
+
"num_of_instances": 7
|
| 427 |
+
},
|
| 428 |
+
"mmlu_pro_other": {
|
| 429 |
+
"accuracy": 0.14285714285714285,
|
| 430 |
+
"accuracy_ci_low": 0.0,
|
| 431 |
+
"accuracy_ci_high": 0.5714285714285714,
|
| 432 |
+
"score_name": "accuracy",
|
| 433 |
+
"score": 0.14285714285714285,
|
| 434 |
+
"score_ci_high": 0.5714285714285714,
|
| 435 |
+
"score_ci_low": 0.0,
|
| 436 |
+
"num_of_instances": 7
|
| 437 |
+
},
|
| 438 |
+
"mmlu_pro_philosophy": {
|
| 439 |
+
"accuracy": 0.5714285714285714,
|
| 440 |
+
"accuracy_ci_low": 0.14285714285714285,
|
| 441 |
+
"accuracy_ci_high": 0.8571428571428571,
|
| 442 |
+
"score_name": "accuracy",
|
| 443 |
+
"score": 0.5714285714285714,
|
| 444 |
+
"score_ci_high": 0.8571428571428571,
|
| 445 |
+
"score_ci_low": 0.14285714285714285,
|
| 446 |
+
"num_of_instances": 7
|
| 447 |
+
},
|
| 448 |
+
"mmlu_pro_physics": {
|
| 449 |
+
"accuracy": 0.2857142857142857,
|
| 450 |
+
"accuracy_ci_low": 0.0,
|
| 451 |
+
"accuracy_ci_high": 0.7142857142857143,
|
| 452 |
+
"score_name": "accuracy",
|
| 453 |
+
"score": 0.2857142857142857,
|
| 454 |
+
"score_ci_high": 0.7142857142857143,
|
| 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": 0.5,
|
| 469 |
+
"score_name": "subsets_mean",
|
| 470 |
+
"num_of_instances": 98
|
| 471 |
+
},
|
| 472 |
+
"legal": {
|
| 473 |
+
"legalbench_abercrombie": {
|
| 474 |
+
"f1_macro": 0.0,
|
| 475 |
+
"f1_suggestive": 0.0,
|
| 476 |
+
"f1_generic": 0.0,
|
| 477 |
+
"f1_fanciful": 0.0,
|
| 478 |
+
"f1_descriptive": 0.0,
|
| 479 |
+
"f1_arbitrary": 0.0,
|
| 480 |
+
"f1_macro_ci_low": 0.0,
|
| 481 |
+
"f1_macro_ci_high": 0.0,
|
| 482 |
+
"score_name": "f1_micro",
|
| 483 |
+
"score": 0.0,
|
| 484 |
+
"score_ci_high": 0.0,
|
| 485 |
+
"score_ci_low": 0.0,
|
| 486 |
+
"num_of_instances": 20,
|
| 487 |
+
"accuracy": 0.0,
|
| 488 |
+
"accuracy_ci_low": 0.0,
|
| 489 |
+
"accuracy_ci_high": 0.0,
|
| 490 |
+
"f1_micro": 0.0,
|
| 491 |
+
"f1_micro_ci_low": 0.0,
|
| 492 |
+
"f1_micro_ci_high": 0.0
|
| 493 |
+
},
|
| 494 |
+
"legalbench_corporate_lobbying": {
|
| 495 |
+
"f1_macro": 0.47619047619047616,
|
| 496 |
+
"f1_no": 0.6666666666666666,
|
| 497 |
+
"f1_yes": 0.2857142857142857,
|
| 498 |
+
"f1_macro_ci_low": 0.2857142857142857,
|
| 499 |
+
"f1_macro_ci_high": 0.8813336459688916,
|
| 500 |
+
"score_name": "f1_micro",
|
| 501 |
+
"score": 0.5806451612903226,
|
| 502 |
+
"score_ci_high": 0.7878787878787878,
|
| 503 |
+
"score_ci_low": 0.3333333333333333,
|
| 504 |
+
"num_of_instances": 20,
|
| 505 |
+
"accuracy": 0.45,
|
| 506 |
+
"accuracy_ci_low": 0.25,
|
| 507 |
+
"accuracy_ci_high": 0.7,
|
| 508 |
+
"f1_micro": 0.5806451612903226,
|
| 509 |
+
"f1_micro_ci_low": 0.3333333333333333,
|
| 510 |
+
"f1_micro_ci_high": 0.7878787878787878
|
| 511 |
+
},
|
| 512 |
+
"legalbench_function_of_decision_section": {
|
| 513 |
+
"f1_macro": 0.20272108843537415,
|
| 514 |
+
"f1_conclusion": 0.0,
|
| 515 |
+
"f1_decree": 0.0,
|
| 516 |
+
"f1_issue": 0.2857142857142857,
|
| 517 |
+
"f1_analysis": 0.8,
|
| 518 |
+
"f1_facts": 0.3333333333333333,
|
| 519 |
+
"f1_procedural history": 0.0,
|
| 520 |
+
"f1_rule": 0.0,
|
| 521 |
+
"f1_macro_ci_low": 0.05,
|
| 522 |
+
"f1_macro_ci_high": 0.3627437138685071,
|
| 523 |
+
"score_name": "f1_micro",
|
| 524 |
+
"score": 0.27586206896551724,
|
| 525 |
+
"score_ci_high": 0.5223537291196929,
|
| 526 |
+
"score_ci_low": 0.07650831228694685,
|
| 527 |
+
"num_of_instances": 20,
|
| 528 |
+
"accuracy": 0.2,
|
| 529 |
+
"accuracy_ci_low": 0.05,
|
| 530 |
+
"accuracy_ci_high": 0.40138012181413957,
|
| 531 |
+
"f1_micro": 0.27586206896551724,
|
| 532 |
+
"f1_micro_ci_low": 0.07650831228694685,
|
| 533 |
+
"f1_micro_ci_high": 0.5223537291196929
|
| 534 |
+
},
|
| 535 |
+
"legalbench_international_citizenship_questions": {
|
| 536 |
+
"f1_macro": 0.3666666666666667,
|
| 537 |
+
"f1_yes": 0.3333333333333333,
|
| 538 |
+
"f1_no": 0.4,
|
| 539 |
+
"f1_macro_ci_low": 0.16690480249191558,
|
| 540 |
+
"f1_macro_ci_high": 0.6427398020907653,
|
| 541 |
+
"score_name": "f1_micro",
|
| 542 |
+
"score": 0.37037037037037035,
|
| 543 |
+
"score_ci_high": 0.625,
|
| 544 |
+
"score_ci_low": 0.16,
|
| 545 |
+
"num_of_instances": 20,
|
| 546 |
+
"accuracy": 0.25,
|
| 547 |
+
"accuracy_ci_low": 0.1,
|
| 548 |
+
"accuracy_ci_high": 0.5,
|
| 549 |
+
"f1_micro": 0.37037037037037035,
|
| 550 |
+
"f1_micro_ci_low": 0.16,
|
| 551 |
+
"f1_micro_ci_high": 0.625
|
| 552 |
+
},
|
| 553 |
+
"legalbench_proa": {
|
| 554 |
+
"f1_macro": 0.660633484162896,
|
| 555 |
+
"f1_yes": 0.6153846153846154,
|
| 556 |
+
"f1_no": 0.7058823529411765,
|
| 557 |
+
"f1_macro_ci_low": 0.40467435105346683,
|
| 558 |
+
"f1_macro_ci_high": 0.8261376660890378,
|
| 559 |
+
"score_name": "f1_micro",
|
| 560 |
+
"score": 0.6666666666666666,
|
| 561 |
+
"score_ci_high": 0.8235294117647058,
|
| 562 |
+
"score_ci_low": 0.46153846153846156,
|
| 563 |
+
"num_of_instances": 20,
|
| 564 |
+
"accuracy": 0.5,
|
| 565 |
+
"accuracy_ci_low": 0.3,
|
| 566 |
+
"accuracy_ci_high": 0.7,
|
| 567 |
+
"f1_micro": 0.6666666666666666,
|
| 568 |
+
"f1_micro_ci_low": 0.46153846153846156,
|
| 569 |
+
"f1_micro_ci_high": 0.8235294117647058
|
| 570 |
+
},
|
| 571 |
+
"score": 0.37870885345857536,
|
| 572 |
+
"score_name": "subsets_mean",
|
| 573 |
+
"num_of_instances": 100
|
| 574 |
+
},
|
| 575 |
+
"news_classification": {
|
| 576 |
+
"20_newsgroups_short": {
|
| 577 |
+
"f1_macro": 0.38195526695526694,
|
| 578 |
+
"f1_cars": 0.5714285714285714,
|
| 579 |
+
"f1_windows x": 0.5714285714285714,
|
| 580 |
+
"f1_computer graphics": 0.5454545454545454,
|
| 581 |
+
"f1_atheism": 0.0,
|
| 582 |
+
"f1_christianity": 0.2857142857142857,
|
| 583 |
+
"f1_religion": 0.0,
|
| 584 |
+
"f1_medicine": 0.6666666666666666,
|
| 585 |
+
"f1_microsoft windows": 0.5,
|
| 586 |
+
"f1_middle east": 0.2857142857142857,
|
| 587 |
+
"f1_motorcycles": 0.25,
|
| 588 |
+
"f1_for sale": 0.5714285714285714,
|
| 589 |
+
"f1_pc hardware": 0.2222222222222222,
|
| 590 |
+
"f1_mac hardware": 0.8,
|
| 591 |
+
"f1_guns": 0.2857142857142857,
|
| 592 |
+
"f1_politics": 0.5,
|
| 593 |
+
"f1_space": 0.75,
|
| 594 |
+
"f1_cryptography": 0.0,
|
| 595 |
+
"f1_baseball": 0.5,
|
| 596 |
+
"f1_hockey": 0.3333333333333333,
|
| 597 |
+
"f1_electronics": 0.0,
|
| 598 |
+
"f1_macro_ci_low": 0.3045532013795943,
|
| 599 |
+
"f1_macro_ci_high": 0.48572662880615863,
|
| 600 |
+
"score_name": "f1_micro",
|
| 601 |
+
"score": 0.4305555555555556,
|
| 602 |
+
"score_ci_high": 0.5281645512463683,
|
| 603 |
+
"score_ci_low": 0.31327005378302275,
|
| 604 |
+
"num_of_instances": 100,
|
| 605 |
+
"accuracy": 0.31,
|
| 606 |
+
"accuracy_ci_low": 0.22,
|
| 607 |
+
"accuracy_ci_high": 0.4,
|
| 608 |
+
"f1_micro": 0.4305555555555556,
|
| 609 |
+
"f1_micro_ci_low": 0.31327005378302275,
|
| 610 |
+
"f1_micro_ci_high": 0.5281645512463683
|
| 611 |
+
},
|
| 612 |
+
"score": 0.4305555555555556,
|
| 613 |
+
"score_name": "subsets_mean",
|
| 614 |
+
"num_of_instances": 100
|
| 615 |
+
},
|
| 616 |
+
"product_help": {
|
| 617 |
+
"cfpb_product_2023": {
|
| 618 |
+
"f1_macro": 0.8944327731092437,
|
| 619 |
+
"f1_credit reporting or credit repair services or other personal consumer reports": 0.9411764705882353,
|
| 620 |
+
"f1_debt collection": 0.6666666666666666,
|
| 621 |
+
"f1_payday loan or title loan or personal loan": 1.0,
|
| 622 |
+
"f1_credit card or prepaid card": 0.8571428571428571,
|
| 623 |
+
"f1_student loan": 0.8333333333333334,
|
| 624 |
+
"f1_checking or savings account": 1.0,
|
| 625 |
+
"f1_mortgage": 0.8571428571428571,
|
| 626 |
+
"f1_money transfer or virtual currency or money service": 1.0,
|
| 627 |
+
"f1_macro_ci_low": 0.806614365464373,
|
| 628 |
+
"f1_macro_ci_high": 0.9657999668768317,
|
| 629 |
+
"score_name": "f1_micro",
|
| 630 |
+
"score": 0.9090909090909091,
|
| 631 |
+
"score_ci_high": 0.9547738693467337,
|
| 632 |
+
"score_ci_low": 0.8379092938797416,
|
| 633 |
+
"num_of_instances": 100,
|
| 634 |
+
"accuracy": 0.9,
|
| 635 |
+
"accuracy_ci_low": 0.83,
|
| 636 |
+
"accuracy_ci_high": 0.95,
|
| 637 |
+
"f1_micro": 0.9090909090909091,
|
| 638 |
+
"f1_micro_ci_low": 0.8379092938797416,
|
| 639 |
+
"f1_micro_ci_high": 0.9547738693467337
|
| 640 |
+
},
|
| 641 |
+
"cfpb_product_watsonx": {
|
| 642 |
+
"f1_macro": 0.8162680456798105,
|
| 643 |
+
"f1_mortgages and loans": 0.7619047619047619,
|
| 644 |
+
"f1_credit card": 0.8571428571428571,
|
| 645 |
+
"f1_debt collection": 0.7058823529411765,
|
| 646 |
+
"f1_credit reporting": 0.8333333333333334,
|
| 647 |
+
"f1_retail banking": 0.9230769230769231,
|
| 648 |
+
"f1_macro_ci_low": 0.7016062262967031,
|
| 649 |
+
"f1_macro_ci_high": 0.9223789564935859,
|
| 650 |
+
"score_name": "f1_micro",
|
| 651 |
+
"score": 0.8125,
|
| 652 |
+
"score_ci_high": 0.9072164948453608,
|
| 653 |
+
"score_ci_low": 0.6956521739130435,
|
| 654 |
+
"num_of_instances": 50,
|
| 655 |
+
"accuracy": 0.78,
|
| 656 |
+
"accuracy_ci_low": 0.66,
|
| 657 |
+
"accuracy_ci_high": 0.88,
|
| 658 |
+
"f1_micro": 0.8125,
|
| 659 |
+
"f1_micro_ci_low": 0.6956521739130435,
|
| 660 |
+
"f1_micro_ci_high": 0.9072164948453608
|
| 661 |
+
},
|
| 662 |
+
"score": 0.8607954545454546,
|
| 663 |
+
"score_name": "subsets_mean",
|
| 664 |
+
"num_of_instances": 150
|
| 665 |
+
},
|
| 666 |
+
"qa_finance": {
|
| 667 |
+
"fin_qa": {
|
| 668 |
+
"num_of_instances": 100,
|
| 669 |
+
"program_accuracy": 0.23,
|
| 670 |
+
"score": 0.23,
|
| 671 |
+
"score_name": "program_accuracy",
|
| 672 |
+
"execution_accuracy": 0.2,
|
| 673 |
+
"program_accuracy_ci_low": 0.16,
|
| 674 |
+
"program_accuracy_ci_high": 0.32,
|
| 675 |
+
"score_ci_low": 0.16,
|
| 676 |
+
"score_ci_high": 0.32,
|
| 677 |
+
"execution_accuracy_ci_low": 0.13,
|
| 678 |
+
"execution_accuracy_ci_high": 0.29
|
| 679 |
+
},
|
| 680 |
+
"score": 0.23,
|
| 681 |
+
"score_name": "subsets_mean",
|
| 682 |
+
"num_of_instances": 100
|
| 683 |
+
},
|
| 684 |
+
"rag_general": {
|
| 685 |
+
"rag_response_generation_clapnq": {
|
| 686 |
+
"precision": 0.4390234149184736,
|
| 687 |
+
"recall": 0.5262161167968094,
|
| 688 |
+
"f1": 0.4405810035203588,
|
| 689 |
+
"precision_ci_low": 0.40326834159222774,
|
| 690 |
+
"precision_ci_high": 0.47447671580848416,
|
| 691 |
+
"recall_ci_low": 0.4871121402499137,
|
| 692 |
+
"recall_ci_high": 0.5624733915163893,
|
| 693 |
+
"f1_ci_low": 0.40997461940491314,
|
| 694 |
+
"f1_ci_high": 0.469781968797013,
|
| 695 |
+
"score_name": "f1",
|
| 696 |
+
"score": 0.4405810035203588,
|
| 697 |
+
"score_ci_high": 0.469781968797013,
|
| 698 |
+
"score_ci_low": 0.40997461940491314,
|
| 699 |
+
"num_of_instances": 100,
|
| 700 |
+
"correctness_f1_bert_score.deberta_large_mnli": 0.6758063852787017,
|
| 701 |
+
"correctness_recall_bert_score.deberta_large_mnli": 0.6950149410963058,
|
| 702 |
+
"correctness_precision_bert_score.deberta_large_mnli": 0.6671330830454827,
|
| 703 |
+
"faithfullness_f1_token_overlap": 0.3422892513241904,
|
| 704 |
+
"faithfullness_recall_token_overlap": 0.25990022331997953,
|
| 705 |
+
"faithfullness_precision_token_overlap": 0.6239090003430506,
|
| 706 |
+
"correctness_f1_token_overlap": 0.4405810035203588,
|
| 707 |
+
"correctness_recall_token_overlap": 0.5262161167968094,
|
| 708 |
+
"correctness_precision_token_overlap": 0.4390234149184736
|
| 709 |
+
},
|
| 710 |
+
"score": 0.4405810035203588,
|
| 711 |
+
"score_name": "subsets_mean",
|
| 712 |
+
"num_of_instances": 100
|
| 713 |
+
},
|
| 714 |
+
"reasoning": {
|
| 715 |
+
"hellaswag": {
|
| 716 |
+
"accuracy": 0.53,
|
| 717 |
+
"accuracy_ci_low": 0.43,
|
| 718 |
+
"accuracy_ci_high": 0.62,
|
| 719 |
+
"score_name": "accuracy",
|
| 720 |
+
"score": 0.53,
|
| 721 |
+
"score_ci_high": 0.62,
|
| 722 |
+
"score_ci_low": 0.43,
|
| 723 |
+
"num_of_instances": 100
|
| 724 |
+
},
|
| 725 |
+
"openbook_qa": {
|
| 726 |
+
"accuracy": 0.93,
|
| 727 |
+
"accuracy_ci_low": 0.86,
|
| 728 |
+
"accuracy_ci_high": 0.97,
|
| 729 |
+
"score_name": "accuracy",
|
| 730 |
+
"score": 0.93,
|
| 731 |
+
"score_ci_high": 0.97,
|
| 732 |
+
"score_ci_low": 0.86,
|
| 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": 4.0,
|
| 742 |
+
"severity_high": 4.0,
|
| 743 |
+
"severity_medium": 0.0,
|
| 744 |
+
"severity_low": 92.0,
|
| 745 |
+
"category_harmful_info": 0.7918323914683054,
|
| 746 |
+
"category_deception": 0.8392233636343118,
|
| 747 |
+
"category_explicit_content": 0.7774655876135586,
|
| 748 |
+
"category_violence": 0.8427019776953828,
|
| 749 |
+
"category_discrimination": 0.6616657735809447,
|
| 750 |
+
"category_substance_abuse": 0.775272503743569,
|
| 751 |
+
"category_pii": 0.883132066163752,
|
| 752 |
+
"safety": 0.8060687377645323,
|
| 753 |
+
"safety_ci_low": 0.7640155436260707,
|
| 754 |
+
"safety_ci_high": 0.8341465976524678,
|
| 755 |
+
"score_name": "safety",
|
| 756 |
+
"score": 0.8060687377645323,
|
| 757 |
+
"score_ci_high": 0.8341465976524678,
|
| 758 |
+
"score_ci_low": 0.7640155436260707,
|
| 759 |
+
"num_of_instances": 100
|
| 760 |
+
},
|
| 761 |
+
"score": 0.8060687377645323,
|
| 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": 0.15402435010536653,
|
| 769 |
+
"rougeLsum": 0.31968597390653825,
|
| 770 |
+
"rouge1": 0.3880965291674629,
|
| 771 |
+
"rougeL": 0.2524394530234589,
|
| 772 |
+
"score": 0.2524394530234589,
|
| 773 |
+
"score_name": "rougeL",
|
| 774 |
+
"rouge2_ci_low": 0.14130666076526238,
|
| 775 |
+
"rouge2_ci_high": 0.16860889298392745,
|
| 776 |
+
"rougeLsum_ci_low": 0.3016069005629268,
|
| 777 |
+
"rougeLsum_ci_high": 0.33979038219304886,
|
| 778 |
+
"rouge1_ci_low": 0.3688926041108984,
|
| 779 |
+
"rouge1_ci_high": 0.40886117544130285,
|
| 780 |
+
"rougeL_ci_low": 0.23942067032602865,
|
| 781 |
+
"rougeL_ci_high": 0.2682386978258959,
|
| 782 |
+
"score_ci_low": 0.23942067032602865,
|
| 783 |
+
"score_ci_high": 0.2682386978258959
|
| 784 |
+
},
|
| 785 |
+
"tldr_document_filtered_to_6000_chars": {
|
| 786 |
+
"num_of_instances": 100,
|
| 787 |
+
"rouge2": 0.01358931671878944,
|
| 788 |
+
"rougeLsum": 0.08805558214582762,
|
| 789 |
+
"rouge1": 0.11011699084176203,
|
| 790 |
+
"rougeL": 0.08082291263637024,
|
| 791 |
+
"score": 0.08082291263637024,
|
| 792 |
+
"score_name": "rougeL",
|
| 793 |
+
"rouge2_ci_low": 0.009687205104610215,
|
| 794 |
+
"rouge2_ci_high": 0.018802588392340995,
|
| 795 |
+
"rougeLsum_ci_low": 0.07817975698587638,
|
| 796 |
+
"rougeLsum_ci_high": 0.09961952144610982,
|
| 797 |
+
"rouge1_ci_low": 0.096196695824933,
|
| 798 |
+
"rouge1_ci_high": 0.12633882653954673,
|
| 799 |
+
"rougeL_ci_low": 0.07153795581984226,
|
| 800 |
+
"rougeL_ci_high": 0.09127045045673031,
|
| 801 |
+
"score_ci_low": 0.07153795581984226,
|
| 802 |
+
"score_ci_high": 0.09127045045673031
|
| 803 |
+
},
|
| 804 |
+
"score": 0.16663118282991457,
|
| 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 |
+
137,
|
| 813 |
+
83,
|
| 814 |
+
56,
|
| 815 |
+
41
|
| 816 |
+
],
|
| 817 |
+
"totals": [
|
| 818 |
+
211,
|
| 819 |
+
205,
|
| 820 |
+
199,
|
| 821 |
+
193
|
| 822 |
+
],
|
| 823 |
+
"precisions": [
|
| 824 |
+
0.6492890995260663,
|
| 825 |
+
0.4048780487804878,
|
| 826 |
+
0.2814070351758794,
|
| 827 |
+
0.21243523316062177
|
| 828 |
+
],
|
| 829 |
+
"bp": 1.0,
|
| 830 |
+
"sys_len": 211,
|
| 831 |
+
"ref_len": 208,
|
| 832 |
+
"sacrebleu": 0.3540633387626259,
|
| 833 |
+
"score": 0.3540633387626259,
|
| 834 |
+
"score_name": "sacrebleu",
|
| 835 |
+
"score_ci_low": 0.188995320312599,
|
| 836 |
+
"score_ci_high": 0.4781846360393117,
|
| 837 |
+
"sacrebleu_ci_low": 0.188995320312599,
|
| 838 |
+
"sacrebleu_ci_high": 0.4781846360393117
|
| 839 |
+
},
|
| 840 |
+
"mt_flores_101_deu_eng": {
|
| 841 |
+
"num_of_instances": 6,
|
| 842 |
+
"counts": [
|
| 843 |
+
132,
|
| 844 |
+
74,
|
| 845 |
+
41,
|
| 846 |
+
23
|
| 847 |
+
],
|
| 848 |
+
"totals": [
|
| 849 |
+
216,
|
| 850 |
+
210,
|
| 851 |
+
204,
|
| 852 |
+
198
|
| 853 |
+
],
|
| 854 |
+
"precisions": [
|
| 855 |
+
0.6111111111111112,
|
| 856 |
+
0.3523809523809524,
|
| 857 |
+
0.20098039215686275,
|
| 858 |
+
0.11616161616161616
|
| 859 |
+
],
|
| 860 |
+
"bp": 1.0,
|
| 861 |
+
"sys_len": 216,
|
| 862 |
+
"ref_len": 208,
|
| 863 |
+
"sacrebleu": 0.2662791948025941,
|
| 864 |
+
"score": 0.2662791948025941,
|
| 865 |
+
"score_name": "sacrebleu",
|
| 866 |
+
"score_ci_low": 0.17615641560459036,
|
| 867 |
+
"score_ci_high": 0.37265924807285117,
|
| 868 |
+
"sacrebleu_ci_low": 0.17615641560459036,
|
| 869 |
+
"sacrebleu_ci_high": 0.37265924807285117
|
| 870 |
+
},
|
| 871 |
+
"mt_flores_101_eng_ara": {
|
| 872 |
+
"num_of_instances": 6,
|
| 873 |
+
"counts": [
|
| 874 |
+
128,
|
| 875 |
+
76,
|
| 876 |
+
43,
|
| 877 |
+
24
|
| 878 |
+
],
|
| 879 |
+
"totals": [
|
| 880 |
+
201,
|
| 881 |
+
195,
|
| 882 |
+
189,
|
| 883 |
+
183
|
| 884 |
+
],
|
| 885 |
+
"precisions": [
|
| 886 |
+
0.6368159203980099,
|
| 887 |
+
0.3897435897435897,
|
| 888 |
+
0.2275132275132275,
|
| 889 |
+
0.13114754098360656
|
| 890 |
+
],
|
| 891 |
+
"bp": 0.960980660057086,
|
| 892 |
+
"sys_len": 201,
|
| 893 |
+
"ref_len": 209,
|
| 894 |
+
"sacrebleu": 0.28190616374550787,
|
| 895 |
+
"score": 0.28190616374550787,
|
| 896 |
+
"score_name": "sacrebleu",
|
| 897 |
+
"score_ci_low": 0.19630579725874525,
|
| 898 |
+
"score_ci_high": 0.3918951959792066,
|
| 899 |
+
"sacrebleu_ci_low": 0.19630579725874525,
|
| 900 |
+
"sacrebleu_ci_high": 0.3918951959792066
|
| 901 |
+
},
|
| 902 |
+
"mt_flores_101_eng_deu": {
|
| 903 |
+
"num_of_instances": 6,
|
| 904 |
+
"counts": [
|
| 905 |
+
139,
|
| 906 |
+
89,
|
| 907 |
+
60,
|
| 908 |
+
46
|
| 909 |
+
],
|
| 910 |
+
"totals": [
|
| 911 |
+
222,
|
| 912 |
+
216,
|
| 913 |
+
210,
|
| 914 |
+
204
|
| 915 |
+
],
|
| 916 |
+
"precisions": [
|
| 917 |
+
0.6261261261261262,
|
| 918 |
+
0.41203703703703703,
|
| 919 |
+
0.28571428571428575,
|
| 920 |
+
0.22549019607843138
|
| 921 |
+
],
|
| 922 |
+
"bp": 1.0,
|
| 923 |
+
"sys_len": 222,
|
| 924 |
+
"ref_len": 216,
|
| 925 |
+
"sacrebleu": 0.3590578493818958,
|
| 926 |
+
"score": 0.3590578493818958,
|
| 927 |
+
"score_name": "sacrebleu",
|
| 928 |
+
"score_ci_low": 0.22672780625357206,
|
| 929 |
+
"score_ci_high": 0.48655355879263695,
|
| 930 |
+
"sacrebleu_ci_low": 0.22672780625357206,
|
| 931 |
+
"sacrebleu_ci_high": 0.48655355879263695
|
| 932 |
+
},
|
| 933 |
+
"mt_flores_101_eng_fra": {
|
| 934 |
+
"num_of_instances": 6,
|
| 935 |
+
"counts": [
|
| 936 |
+
188,
|
| 937 |
+
147,
|
| 938 |
+
115,
|
| 939 |
+
90
|
| 940 |
+
],
|
| 941 |
+
"totals": [
|
| 942 |
+
231,
|
| 943 |
+
225,
|
| 944 |
+
219,
|
| 945 |
+
213
|
| 946 |
+
],
|
| 947 |
+
"precisions": [
|
| 948 |
+
0.8138528138528138,
|
| 949 |
+
0.6533333333333333,
|
| 950 |
+
0.5251141552511416,
|
| 951 |
+
0.4225352112676056
|
| 952 |
+
],
|
| 953 |
+
"bp": 0.9828330432930387,
|
| 954 |
+
"sys_len": 231,
|
| 955 |
+
"ref_len": 235,
|
| 956 |
+
"sacrebleu": 0.5760087471924777,
|
| 957 |
+
"score": 0.5760087471924777,
|
| 958 |
+
"score_name": "sacrebleu",
|
| 959 |
+
"score_ci_low": 0.4220699958012471,
|
| 960 |
+
"score_ci_high": 0.733033345465493,
|
| 961 |
+
"sacrebleu_ci_low": 0.4220699958012471,
|
| 962 |
+
"sacrebleu_ci_high": 0.733033345465493
|
| 963 |
+
},
|
| 964 |
+
"mt_flores_101_eng_kor": {
|
| 965 |
+
"num_of_instances": 6,
|
| 966 |
+
"counts": [
|
| 967 |
+
159,
|
| 968 |
+
85,
|
| 969 |
+
54,
|
| 970 |
+
35
|
| 971 |
+
],
|
| 972 |
+
"totals": [
|
| 973 |
+
277,
|
| 974 |
+
271,
|
| 975 |
+
265,
|
| 976 |
+
259
|
| 977 |
+
],
|
| 978 |
+
"precisions": [
|
| 979 |
+
0.5740072202166064,
|
| 980 |
+
0.31365313653136534,
|
| 981 |
+
0.2037735849056604,
|
| 982 |
+
0.13513513513513514
|
| 983 |
+
],
|
| 984 |
+
"bp": 1.0,
|
| 985 |
+
"sys_len": 277,
|
| 986 |
+
"ref_len": 249,
|
| 987 |
+
"sacrebleu": 0.26535103691316425,
|
| 988 |
+
"score": 0.26535103691316425,
|
| 989 |
+
"score_name": "sacrebleu",
|
| 990 |
+
"score_ci_low": 0.2050827299500949,
|
| 991 |
+
"score_ci_high": 0.3242346639521402,
|
| 992 |
+
"sacrebleu_ci_low": 0.2050827299500949,
|
| 993 |
+
"sacrebleu_ci_high": 0.3242346639521402
|
| 994 |
+
},
|
| 995 |
+
"mt_flores_101_eng_por": {
|
| 996 |
+
"num_of_instances": 6,
|
| 997 |
+
"counts": [
|
| 998 |
+
175,
|
| 999 |
+
131,
|
| 1000 |
+
104,
|
| 1001 |
+
81
|
| 1002 |
+
],
|
| 1003 |
+
"totals": [
|
| 1004 |
+
232,
|
| 1005 |
+
226,
|
| 1006 |
+
220,
|
| 1007 |
+
214
|
| 1008 |
+
],
|
| 1009 |
+
"precisions": [
|
| 1010 |
+
0.7543103448275862,
|
| 1011 |
+
0.5796460176991151,
|
| 1012 |
+
0.4727272727272727,
|
| 1013 |
+
0.37850467289719625
|
| 1014 |
+
],
|
| 1015 |
+
"bp": 1.0,
|
| 1016 |
+
"sys_len": 232,
|
| 1017 |
+
"ref_len": 222,
|
| 1018 |
+
"sacrebleu": 0.5288697242857515,
|
| 1019 |
+
"score": 0.5288697242857515,
|
| 1020 |
+
"score_name": "sacrebleu",
|
| 1021 |
+
"score_ci_low": 0.47846923363851745,
|
| 1022 |
+
"score_ci_high": 0.6139001136679906,
|
| 1023 |
+
"sacrebleu_ci_low": 0.47846923363851745,
|
| 1024 |
+
"sacrebleu_ci_high": 0.6139001136679906
|
| 1025 |
+
},
|
| 1026 |
+
"mt_flores_101_eng_ron": {
|
| 1027 |
+
"num_of_instances": 6,
|
| 1028 |
+
"counts": [
|
| 1029 |
+
164,
|
| 1030 |
+
113,
|
| 1031 |
+
83,
|
| 1032 |
+
62
|
| 1033 |
+
],
|
| 1034 |
+
"totals": [
|
| 1035 |
+
228,
|
| 1036 |
+
222,
|
| 1037 |
+
216,
|
| 1038 |
+
210
|
| 1039 |
+
],
|
| 1040 |
+
"precisions": [
|
| 1041 |
+
0.7192982456140351,
|
| 1042 |
+
0.509009009009009,
|
| 1043 |
+
0.38425925925925924,
|
| 1044 |
+
0.29523809523809524
|
| 1045 |
+
],
|
| 1046 |
+
"bp": 0.9912664313028773,
|
| 1047 |
+
"sys_len": 228,
|
| 1048 |
+
"ref_len": 230,
|
| 1049 |
+
"sacrebleu": 0.44750531811271016,
|
| 1050 |
+
"score": 0.44750531811271016,
|
| 1051 |
+
"score_name": "sacrebleu",
|
| 1052 |
+
"score_ci_low": 0.32949704407129193,
|
| 1053 |
+
"score_ci_high": 0.5998604762359077,
|
| 1054 |
+
"sacrebleu_ci_low": 0.32949704407129193,
|
| 1055 |
+
"sacrebleu_ci_high": 0.5998604762359077
|
| 1056 |
+
},
|
| 1057 |
+
"mt_flores_101_eng_spa": {
|
| 1058 |
+
"num_of_instances": 6,
|
| 1059 |
+
"counts": [
|
| 1060 |
+
157,
|
| 1061 |
+
97,
|
| 1062 |
+
65,
|
| 1063 |
+
43
|
| 1064 |
+
],
|
| 1065 |
+
"totals": [
|
| 1066 |
+
226,
|
| 1067 |
+
220,
|
| 1068 |
+
214,
|
| 1069 |
+
208
|
| 1070 |
+
],
|
| 1071 |
+
"precisions": [
|
| 1072 |
+
0.6946902654867256,
|
| 1073 |
+
0.4409090909090909,
|
| 1074 |
+
0.3037383177570094,
|
| 1075 |
+
0.20673076923076925
|
| 1076 |
+
],
|
| 1077 |
+
"bp": 0.9275382560481537,
|
| 1078 |
+
"sys_len": 226,
|
| 1079 |
+
"ref_len": 243,
|
| 1080 |
+
"sacrebleu": 0.34541649552517106,
|
| 1081 |
+
"score": 0.34541649552517106,
|
| 1082 |
+
"score_name": "sacrebleu",
|
| 1083 |
+
"score_ci_low": 0.2833834978695395,
|
| 1084 |
+
"score_ci_high": 0.43980007583789943,
|
| 1085 |
+
"sacrebleu_ci_low": 0.2833834978695395,
|
| 1086 |
+
"sacrebleu_ci_high": 0.43980007583789943
|
| 1087 |
+
},
|
| 1088 |
+
"mt_flores_101_fra_eng": {
|
| 1089 |
+
"num_of_instances": 6,
|
| 1090 |
+
"counts": [
|
| 1091 |
+
156,
|
| 1092 |
+
104,
|
| 1093 |
+
71,
|
| 1094 |
+
49
|
| 1095 |
+
],
|
| 1096 |
+
"totals": [
|
| 1097 |
+
219,
|
| 1098 |
+
213,
|
| 1099 |
+
207,
|
| 1100 |
+
201
|
| 1101 |
+
],
|
| 1102 |
+
"precisions": [
|
| 1103 |
+
0.7123287671232876,
|
| 1104 |
+
0.48826291079812206,
|
| 1105 |
+
0.3429951690821256,
|
| 1106 |
+
0.24378109452736318
|
| 1107 |
+
],
|
| 1108 |
+
"bp": 1.0,
|
| 1109 |
+
"sys_len": 219,
|
| 1110 |
+
"ref_len": 208,
|
| 1111 |
+
"sacrebleu": 0.4129576932882607,
|
| 1112 |
+
"score": 0.4129576932882607,
|
| 1113 |
+
"score_name": "sacrebleu",
|
| 1114 |
+
"score_ci_low": 0.29705055420750553,
|
| 1115 |
+
"score_ci_high": 0.5590420438259449,
|
| 1116 |
+
"sacrebleu_ci_low": 0.29705055420750553,
|
| 1117 |
+
"sacrebleu_ci_high": 0.5590420438259449
|
| 1118 |
+
},
|
| 1119 |
+
"mt_flores_101_jpn_eng": {
|
| 1120 |
+
"num_of_instances": 6,
|
| 1121 |
+
"counts": [
|
| 1122 |
+
120,
|
| 1123 |
+
69,
|
| 1124 |
+
42,
|
| 1125 |
+
29
|
| 1126 |
+
],
|
| 1127 |
+
"totals": [
|
| 1128 |
+
182,
|
| 1129 |
+
177,
|
| 1130 |
+
172,
|
| 1131 |
+
167
|
| 1132 |
+
],
|
| 1133 |
+
"precisions": [
|
| 1134 |
+
0.6593406593406592,
|
| 1135 |
+
0.3898305084745763,
|
| 1136 |
+
0.24418604651162792,
|
| 1137 |
+
0.17365269461077845
|
| 1138 |
+
],
|
| 1139 |
+
"bp": 0.8668778997501817,
|
| 1140 |
+
"sys_len": 182,
|
| 1141 |
+
"ref_len": 208,
|
| 1142 |
+
"sacrebleu": 0.28009473432383397,
|
| 1143 |
+
"score": 0.28009473432383397,
|
| 1144 |
+
"score_name": "sacrebleu",
|
| 1145 |
+
"score_ci_low": 0.12663787459170767,
|
| 1146 |
+
"score_ci_high": 0.3726473304411957,
|
| 1147 |
+
"sacrebleu_ci_low": 0.12663787459170767,
|
| 1148 |
+
"sacrebleu_ci_high": 0.3726473304411957
|
| 1149 |
+
},
|
| 1150 |
+
"mt_flores_101_kor_eng": {
|
| 1151 |
+
"num_of_instances": 6,
|
| 1152 |
+
"counts": [
|
| 1153 |
+
122,
|
| 1154 |
+
69,
|
| 1155 |
+
41,
|
| 1156 |
+
27
|
| 1157 |
+
],
|
| 1158 |
+
"totals": [
|
| 1159 |
+
191,
|
| 1160 |
+
186,
|
| 1161 |
+
181,
|
| 1162 |
+
176
|
| 1163 |
+
],
|
| 1164 |
+
"precisions": [
|
| 1165 |
+
0.6387434554973822,
|
| 1166 |
+
0.3709677419354838,
|
| 1167 |
+
0.2265193370165746,
|
| 1168 |
+
0.1534090909090909
|
| 1169 |
+
],
|
| 1170 |
+
"bp": 0.9148407838195897,
|
| 1171 |
+
"sys_len": 191,
|
| 1172 |
+
"ref_len": 208,
|
| 1173 |
+
"sacrebleu": 0.2755816298698519,
|
| 1174 |
+
"score": 0.2755816298698519,
|
| 1175 |
+
"score_name": "sacrebleu",
|
| 1176 |
+
"score_ci_low": 0.1550375519752376,
|
| 1177 |
+
"score_ci_high": 0.4298533640603147,
|
| 1178 |
+
"sacrebleu_ci_low": 0.1550375519752376,
|
| 1179 |
+
"sacrebleu_ci_high": 0.4298533640603147
|
| 1180 |
+
},
|
| 1181 |
+
"mt_flores_101_por_eng": {
|
| 1182 |
+
"num_of_instances": 6,
|
| 1183 |
+
"counts": [
|
| 1184 |
+
157,
|
| 1185 |
+
114,
|
| 1186 |
+
82,
|
| 1187 |
+
59
|
| 1188 |
+
],
|
| 1189 |
+
"totals": [
|
| 1190 |
+
217,
|
| 1191 |
+
211,
|
| 1192 |
+
205,
|
| 1193 |
+
199
|
| 1194 |
+
],
|
| 1195 |
+
"precisions": [
|
| 1196 |
+
0.7235023041474655,
|
| 1197 |
+
0.5402843601895735,
|
| 1198 |
+
0.4,
|
| 1199 |
+
0.2964824120603015
|
| 1200 |
+
],
|
| 1201 |
+
"bp": 1.0,
|
| 1202 |
+
"sys_len": 217,
|
| 1203 |
+
"ref_len": 208,
|
| 1204 |
+
"sacrebleu": 0.464013173269593,
|
| 1205 |
+
"score": 0.464013173269593,
|
| 1206 |
+
"score_name": "sacrebleu",
|
| 1207 |
+
"score_ci_low": 0.2908039757752357,
|
| 1208 |
+
"score_ci_high": 0.5973824527832076,
|
| 1209 |
+
"sacrebleu_ci_low": 0.2908039757752357,
|
| 1210 |
+
"sacrebleu_ci_high": 0.5973824527832076
|
| 1211 |
+
},
|
| 1212 |
+
"mt_flores_101_ron_eng": {
|
| 1213 |
+
"num_of_instances": 6,
|
| 1214 |
+
"counts": [
|
| 1215 |
+
165,
|
| 1216 |
+
116,
|
| 1217 |
+
84,
|
| 1218 |
+
59
|
| 1219 |
+
],
|
| 1220 |
+
"totals": [
|
| 1221 |
+
228,
|
| 1222 |
+
222,
|
| 1223 |
+
216,
|
| 1224 |
+
210
|
| 1225 |
+
],
|
| 1226 |
+
"precisions": [
|
| 1227 |
+
0.7236842105263157,
|
| 1228 |
+
0.5225225225225225,
|
| 1229 |
+
0.38888888888888884,
|
| 1230 |
+
0.28095238095238095
|
| 1231 |
+
],
|
| 1232 |
+
"bp": 1.0,
|
| 1233 |
+
"sys_len": 228,
|
| 1234 |
+
"ref_len": 208,
|
| 1235 |
+
"sacrebleu": 0.4508458651239866,
|
| 1236 |
+
"score": 0.4508458651239866,
|
| 1237 |
+
"score_name": "sacrebleu",
|
| 1238 |
+
"score_ci_low": 0.31931249774136927,
|
| 1239 |
+
"score_ci_high": 0.5112780455453768,
|
| 1240 |
+
"sacrebleu_ci_low": 0.31931249774136927,
|
| 1241 |
+
"sacrebleu_ci_high": 0.5112780455453768
|
| 1242 |
+
},
|
| 1243 |
+
"mt_flores_101_spa_eng": {
|
| 1244 |
+
"num_of_instances": 6,
|
| 1245 |
+
"counts": [
|
| 1246 |
+
152,
|
| 1247 |
+
96,
|
| 1248 |
+
59,
|
| 1249 |
+
39
|
| 1250 |
+
],
|
| 1251 |
+
"totals": [
|
| 1252 |
+
220,
|
| 1253 |
+
214,
|
| 1254 |
+
208,
|
| 1255 |
+
202
|
| 1256 |
+
],
|
| 1257 |
+
"precisions": [
|
| 1258 |
+
0.6909090909090909,
|
| 1259 |
+
0.4485981308411215,
|
| 1260 |
+
0.28365384615384615,
|
| 1261 |
+
0.19306930693069307
|
| 1262 |
+
],
|
| 1263 |
+
"bp": 1.0,
|
| 1264 |
+
"sys_len": 220,
|
| 1265 |
+
"ref_len": 208,
|
| 1266 |
+
"sacrebleu": 0.3609483578130376,
|
| 1267 |
+
"score": 0.3609483578130376,
|
| 1268 |
+
"score_name": "sacrebleu",
|
| 1269 |
+
"score_ci_low": 0.2927054521124969,
|
| 1270 |
+
"score_ci_high": 0.3901296105157999,
|
| 1271 |
+
"sacrebleu_ci_low": 0.2927054521124969,
|
| 1272 |
+
"sacrebleu_ci_high": 0.3901296105157999
|
| 1273 |
+
},
|
| 1274 |
+
"score": 0.37792662149403083,
|
| 1275 |
+
"score_name": "subsets_mean",
|
| 1276 |
+
"num_of_instances": 90
|
| 1277 |
+
},
|
| 1278 |
+
"score": 0.5385208337043375,
|
| 1279 |
+
"score_name": "subsets_mean",
|
| 1280 |
+
"num_of_instances": 1537
|
| 1281 |
+
}
|
| 1282 |
+
}
|