Update README.md
Browse files
README.md
CHANGED
|
@@ -1,3 +1,22 @@
|
|
| 1 |
-
---
|
| 2 |
-
license: apache-2.0
|
| 3 |
-
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: apache-2.0
|
| 3 |
+
---
|
| 4 |
+
|
| 5 |
+
EXL3 quants of [Qwen3-30B-A3B](https://huggingface.co/Qwen/Qwen3-30B-A3B)
|
| 6 |
+
|
| 7 |
+
[2.25 bits per weight](https://huggingface.co/turboderp/Qwen3-30B-A3B-exl3/tree/2.25bpw)
|
| 8 |
+
[3.00 bits per weight](https://huggingface.co/turboderp/Qwen3-30B-A3B-exl3/tree/3.0bpw)
|
| 9 |
+
[4.00 bits per weight](https://huggingface.co/turboderp/Qwen3-30B-A3B-exl3/tree/4.0bpw)
|
| 10 |
+
[5.00 bits per weight](https://huggingface.co/turboderp/Qwen3-30B-A3B-exl3/tree/5.0bpw)
|
| 11 |
+
[6.00 bits per weight](https://huggingface.co/turboderp/Qwen3-30B-A3B-exl3/tree/6.0bpw)
|
| 12 |
+
|
| 13 |
+
While I work out a way to meaningfully measure perplexity for such a sparse model, here are some other tests:
|
| 14 |
+
|
| 15 |
+
| Model | HumanEval pass@1 | KL-div vs FP16 (wiki2 20k tokens) | Top-1 agreement vs FP16 |
|
| 16 |
+
|----------|------------------|-----------------------------------|-------------------------|
|
| 17 |
+
| 2.25 bpw | 88.41% | 0.1416 | 84.78% |
|
| 18 |
+
| 3.00 bpw | 89.63% | 0.0688 | 89.44% |
|
| 19 |
+
| 4.00 bpw | 92.07% | 0.0215 | 94.33% |
|
| 20 |
+
| 5.00 bpw | 93.29% | 0.0094 | 96.24% |
|
| 21 |
+
| 6.00 bpw | 92.68% | 0.0054 | 97.45% |
|
| 22 |
+
| FP16 | 91.46% | - | - |
|