Update README.md

README.md

@@ -32,8 +32,9 @@ base_model: meta-llama/Meta-Llama-3.1-405B-Instruct
 - **License(s):** Llama3.1
 - **Model Developers:** Neural Magic
 
-Quantized version of [Meta-Llama-3.1-405B-Instruct](https://huggingface.co/meta-llama/Meta-Llama-3.1-405B-Instruct).
-It achieves scores within 1.3% of the scores of the unquantized model for MMLU, ARC-Challenge, GSM-8k, Hellaswag, Winogrande and TruthfulQA.
+This model is a quantized version of [Meta-Llama-3.1-405B-Instruct](https://huggingface.co/meta-llama/Meta-Llama-3.1-405B-Instruct).
+It was evaluated on a several tasks to assess the its quality in comparison to the unquatized model, including multiple-choice, math reasoning, and open-ended text generation.
+Meta-Llama-3.1-405B-Instruct-FP8-dynamic achieves 95.8% recovery for the Arena-Hard evaluation, 99.3% for OpenLLM v1 (using Meta's prompting when available), 98.4% for OpenLLM v2, 100.1% for HumanEval pass@1, and 100.4% for HumanEval+ pass@1.
 
 ### Model Optimizations
 
@@ -142,18 +143,26 @@ oneshot(
 model.save_pretrained("Meta-Llama-3.1-405B-Instruct-quantized.w8a8")
 ```
 
-
 ## Evaluation
 
-The model was evaluated on MMLU, ARC-Challenge, GSM-8K, Hellaswag, Winogrande and TruthfulQA.
-Evaluation was conducted using the Neural Magic fork of [lm-evaluation-harness](https://github.com/neuralmagic/lm-evaluation-harness/tree/llama_3.1_instruct) (branch llama_3.1_instruct) and the [vLLM](https://docs.vllm.ai/en/stable/) engine.
-This version of the lm-evaluation-harness includes versions of MMLU, ARC-Challenge and GSM-8K that match the prompting style of [Meta-Llama-3.1-Instruct-evals](https://huggingface.co/datasets/meta-llama/Meta-Llama-3.1-405B-Instruct-evals).
+This model was evaluated on the well-known Arena-Hard, OpenLLM v1, OpenLLM v2, HumanEval, and HumanEval+ benchmarks.
+In all cases, model outputs were generated with the [vLLM](https://docs.vllm.ai/en/stable/) engine.
+
+Arena-Hard evaluations were conducted using the [Arena-Hard-Auto](https://github.com/lmarena/arena-hard-auto) repository.
+The model generated a single answer for each prompt form Arena-Hard, and each answer was judged twice by GPT-4.
+We report below the scores obtained in each judgement and the average.
+
+OpenLLM v1 and v2 evaluations were conducted using Neural Magic's fork of [lm-evaluation-harness](https://github.com/neuralmagic/lm-evaluation-harness/tree/llama_3.1_instruct) (branch llama_3.1_instruct).
+This version of the lm-evaluation-harness includes versions of MMLU, ARC-Challenge and GSM-8K that match the prompting style of [Meta-Llama-3.1-Instruct-evals](https://huggingface.co/datasets/meta-llama/Meta-Llama-3.1-70B-Instruct-evals) and a few fixes to OpenLLM v2 tasks.
+
+HumanEval and HumanEval+ evaluations were conducted using Neural Magic's fork of the [EvalPlus](https://github.com/neuralmagic/evalplus) repository.
+
+Detailed model outputs are available as HuggingFace datasets for [Arena-Hard](https://huggingface.co/datasets/neuralmagic/quantized-llama-3.1-arena-hard-evals), [OpenLLM v2](https://huggingface.co/datasets/neuralmagic/quantized-llama-3.1-leaderboard-v2-evals), and [HumanEval](https://huggingface.co/datasets/neuralmagic/quantized-llama-3.1-humaneval-evals).
 
 **Note:** Results have been updated after Meta modified the chat template.
 
 ### Accuracy
 
-#### Open LLM Leaderboard evaluation scores
 <table>
   <tr>
    <td><strong>Benchmark</strong>
@@ -165,12 +174,26 @@ This version of the lm-evaluation-harness includes versions of MMLU, ARC-Challen
    <td><strong>Recovery</strong>
    </td>
   </tr>
+  <tr>
+   <td><strong>Arena Hard</strong>
+   </td>
+   <td>67.4 (67.3 / 67.5)
+   </td>
+   <td>64.6 (64.3 / 64.8)
+   </td>
+   <td>95.8%
+   </td>
+  </tr>
+  <tr>
+   <td><strong>OpenLLM v1</strong>
+   </td>
+  </tr>
   <tr>
    <td>MMLU (5-shot)
    </td>
-   <td>87.38
+   <td>87.4
    </td>
-   <td>87.05
+   <td>87.1
    </td>
    <td>99.6%
    </td>
@@ -178,9 +201,9 @@ This version of the lm-evaluation-harness includes versions of MMLU, ARC-Challen
   <tr>
    <td>ARC Challenge (0-shot)
    </td>
-   <td>94.97
+   <td>95.0
    </td>
-   <td>94.37
+   <td>94.4
    </td>
    <td>99.4%
    </td>
@@ -188,9 +211,9 @@ This version of the lm-evaluation-harness includes versions of MMLU, ARC-Challen
   <tr>
    <td>GSM-8K (CoT, 8-shot, strict-match)
    </td>
-   <td>96.44
+   <td>96.4
    </td>
-   <td>95.45
+   <td>95.5
    </td>
    <td>99.0%
    </td>
@@ -198,9 +221,9 @@ This version of the lm-evaluation-harness includes versions of MMLU, ARC-Challen
   <tr>
    <td>Hellaswag (10-shot)
    </td> 
-   <td>88.33
+   <td>88.3
    </td>
-   <td>88.15
+   <td>88.2
    </td>
    <td>99.8%
    </td>
@@ -208,9 +231,9 @@ This version of the lm-evaluation-harness includes versions of MMLU, ARC-Challen
   <tr>
    <td>Winogrande (5-shot)
    </td>
-   <td>87.21
+   <td>87.2
    </td>
-   <td>86.11
+   <td>86.1
    </td>
    <td>98.7%
    </td>
@@ -218,9 +241,9 @@ This version of the lm-evaluation-harness includes versions of MMLU, ARC-Challen
   <tr>
    <td>TruthfulQA (0-shot)
    </td>
-   <td>64.64
+   <td>64.6
    </td>
-   <td>64.39
+   <td>64.4
    </td>
    <td>99.6%
    </td>
@@ -228,13 +251,111 @@ This version of the lm-evaluation-harness includes versions of MMLU, ARC-Challen
   <tr>
    <td><strong>Average</strong>
    </td>
-   <td><strong>86.75</strong>
+   <td><strong>86.8</strong>
    </td>
-   <td><strong>86.15</strong>
+   <td><strong>86.2</strong>
    </td>
    <td><strong>99.3%</strong>
    </td>
   </tr>
+  <tr>
+   <td><strong>OpenLLM v2</strong>
+   </td>
+  </tr>
+  <tr>
+   <td>MMLU-Pro (5-shot)
+   </td>
+   <td>59.7
+   </td>
+   <td>58.4
+   </td>
+   <td>97.8%
+   </td>
+  </tr>
+  <tr>
+   <td>IFEval (0-shot)
+   </td>
+   <td>87.7
+   </td>
+   <td>87.0
+   </td>
+   <td>99.2%
+   </td>
+  </tr>
+  <tr>
+   <td>BBH (3-shot)
+   </td>
+   <td>67.0
+   </td>
+   <td>66.7
+   </td>
+   <td>99.6%
+   </td>
+  </tr>
+  <tr>
+   <td>Math-|v|-5 (4-shot)
+   </td>
+   <td>39.0
+   </td>
+   <td>35.8
+   </td>
+   <td>91.9%
+   </td>
+  </tr>
+  <tr>
+   <td>GPQA (0-shot)
+   </td>
+   <td>19.5
+   </td>
+   <td>20.4
+   </td>
+   <td>104.5%
+   </td>
+  </tr>
+  <tr>
+   <td>MuSR (0-shot)
+   </td>
+   <td>19.5
+   </td>
+   <td>19.2
+   </td>
+   <td>98.8%
+   </td>
+  </tr>
+  <tr>
+   <td><strong>Average</strong>
+   </td>
+   <td><strong>48.7</strong>
+   </td>
+   <td><strong>47.9</strong>
+   </td>
+   <td><strong>98.4%</strong>
+   </td>
+  </tr>
+  <tr>
+   <td><strong>Coding</strong>
+   </td>
+  </tr>
+  <tr>
+   <td>HumanEval pass@1
+   </td>
+   <td>86.8
+   </td>
+   <td>86.9
+   </td>
+   <td>100.1%
+   </td>
+  </tr>
+  <tr>
+   <td>HumanEval+ pass@1
+   </td>
+   <td>80.1
+   </td>
+   <td>80.4
+   </td>
+   <td>100.4%
+   </td>
+  </tr>
 </table>
 
 ### Reproduction
@@ -315,4 +436,39 @@ lm_eval \
   --tasks truthfulqa \
   --num_fewshot 0 \
   --batch_size auto
-```
+```
+
+#### OpenLLM v2
+```
+lm_eval \
+  --model vllm \
+  --model_args pretrained="neuralmagic/Meta-Llama-3.1-405B-Instruct-quantized.w8a8",dtype=auto,max_model_len=4096,tensor_parallel_size=8,enable_chunked_prefill=True \
+  --apply_chat_template \
+  --fewshot_as_multiturn \
+  --tasks leaderboard \
+  --batch_size auto
+```
+
+#### HumanEval and HumanEval+
+##### Generation
+```
+python3 codegen/generate.py \
+  --model neuralmagic/Meta-Llama-3.1-405B-Instruct-quantized.w8a8 \
+  --bs 16 \
+  --temperature 0.2 \
+  --n_samples 50 \
+  --root "." \
+  --dataset humaneval \
+  --tp 8
+```
+##### Sanitization
+```
+python3 evalplus/sanitize.py \
+  humaneval/neuralmagic--Meta-Llama-3.1-405B-Instruct-quantized.w8a8_vllm_temp_0.2
+```
+##### Evaluation
+```
+evalplus.evaluate \
+  --dataset humaneval \
+  --samples humaneval/neuralmagic--Meta-Llama-3.1-405B-Instruct-quantized.w8a8_vllm_temp_0.2-sanitized
+```