Update README.md
Browse files
README.md
CHANGED
|
@@ -5,6 +5,32 @@ license: apache-2.0
|
|
| 5 |
|
| 6 |
This is GreenBitAI's pretrained **2-bit** LLaMA model with extreme compression yet still strong performance.
|
| 7 |
|
| 8 |
-
There are instruction-tuned LoRA parameters available [here](https://huggingface.co/GreenBitAI/LLaMA-7B-2bit-alpaca).
|
| 9 |
-
|
| 10 |
Please refer to our [Github page](https://github.com/GreenBitAI/low_bit_llama) for the code to run the model and more information.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 5 |
|
| 6 |
This is GreenBitAI's pretrained **2-bit** LLaMA model with extreme compression yet still strong performance.
|
| 7 |
|
|
|
|
|
|
|
| 8 |
Please refer to our [Github page](https://github.com/GreenBitAI/low_bit_llama) for the code to run the model and more information.
|
| 9 |
+
|
| 10 |
+
## Zero-Shot Evaluation
|
| 11 |
+
| Task | Metric | LLaMA-3B q2g32 | LLaMA-3B q2g16 | LLaMA-3B q2g8 | LLaMA-1 7B q2g32 | LLaMA-2 7B q2g32 | LLaMA 3B FP16 | LLaMA-1 7B FP16 |
|
| 12 |
+
|---------------|--------|---------------|---------------|--------------|------------------|------------------|--------------|-----------------|
|
| 13 |
+
| Openbookqa | acc | 0.196 | 0.238 | 0.242 | 0.224 | 0.246 | 0.27 | 0.29 |
|
| 14 |
+
| | ac_norm| 0.332 | 0.358 | 0.362 | 0.388 | 0.376 | 0.4 | 0.41 |
|
| 15 |
+
| arc_challenge | acc | 0.279 | 0.2978 | 0.3148 | 0.3422 | 0.3268 | 0.34 | 0.39 |
|
| 16 |
+
| | ac_norm| 0.2944 | 0.3319 | 0.3345 | 0.3387 | 0.3387 | 0.37 | 0.41 |
|
| 17 |
+
| hellawswag | acc | 0.4238 | 0.444 | 0.462 | 0.4996 | 0.4961 | 0.49 | 0.68 |
|
| 18 |
+
| | ac_norm| 0.5685 | 0.5988 | 0.6242 | 0.6447 | 0.6464 | 0.67 | 0.73 |
|
| 19 |
+
| piqa | acc | 0.7024 | 0.716 | 0.7291 | 0.7476 | 0.7503 | 0.75 | 0.78 |
|
| 20 |
+
| | ac_norm| 0.7116 | 0.7247 | 0.7312 | 0.7443 | 0.7421 | 0.76 | 0.78 |
|
| 21 |
+
| arc_easy | acc | 0.5997 | 0.646 | 0.6528 | 0.6061 | 0.6174 | 0.69 | 0.68 |
|
| 22 |
+
| | ac_norm| 0.5417 | 0.58 | 0.5972 | 0.4566 | 0.4781 | 0.65 | 0.52 |
|
| 23 |
+
| Winogrande | acc | 0.5683 | 0.5888 | 0.6054 | 0.6283 | 0.6298 | 0.62 | 0.68 |
|
| 24 |
+
| boolq | acc | 0.6281 | 0.6636 | 0.6327 | 0.6425 | 0.7061 | 0.68 | 0.75 |
|
| 25 |
+
| truthfulqa_mc | mc1 | 0.2509 | 0.2118 | 0.2252 | 0.224 | 0.2313 | 0.22 | 0.21 |
|
| 26 |
+
| | mc2 | 0.3962 | 0.3501 | 0.3625 | 0.3702 | 0.3854 | 0.35 | 0.34 |
|
| 27 |
+
| anli_r1 | acc | 0.337 | 0.334 | 0.344 | 0.331 | 0.333 | 0.33 | 0.35 |
|
| 28 |
+
| anli_r2 | acc | 0.335 | 0.332 | 0.331 | 0.326 | 0.349 | 0.32 | 0.34 |
|
| 29 |
+
| anli_r3 | acc | 0.3358 | 0.3383 | 0.3425 | 0.3417 | 0.36 | 0.35 | 0.37 |
|
| 30 |
+
| wic | acc | 0.4984 | 0.5094 | 0.4969 | 0.4984 | 0.4953 | 0.48 | 0.5 |
|
| 31 |
+
| rte | acc | 0.5596 | 0.5993 | 0.5632 | 0.639 | 0.6065 | 0.58 | 0.56 |
|
| 32 |
+
| record | f1 | 0.8502 | 0.8625 | 0.8687 | 0.8859 | 0.8872 | 0.88 | 0.91 |
|
| 33 |
+
| | em | 0.8427 | 0.8545 | 0.8612 | 0.8781 | 0.8801 | 0.89 | 0.91 |
|
| 34 |
+
| Average | | 0.4881 | 0.5037 | 0.5087 | 0.5122 | 0.5181 | 0.528 | 0.5519 |
|
| 35 |
+
|
| 36 |
+
|