Differences in the results of the reproduction test on lm-evaluation-harness

#8
by ThreeGold116 - opened

Ouro-1.4B R4 evaluation results between reproduction and paper

  1. reproduction evaluation setting follow the paper

  2. Result Comparsions

    Benchmark Paper Result Reproduction Result
    mmlu 67.35 66.74
    bbh 71.02 60.77
    gsm8k 78.92 60.80
ByteDance org

Sorry for the delay. I recently finished my internship at ByteDance, so I lost control of the repository for a period of time. Regarding the results, the MMLU scores are actually quite consistent (67.35 vs 66.74), likely because the paper reports log-prob results while we used a standard 5-shot setting in lm-eval. For the discrepancies in generate-until tasks like BBH and GSM8K, we used vLLM as the backend to speed up the evaluation since standard generation is quite slow. I suspect the performance drop is due to vLLM-specific behaviors rather than the model itself, so I wanted to ask if the original paper measurements were done without vLLM.

Hi @ThreeGold116 and @ridger ,

The reported performance drops might be due to using the chat template during the standard base evals. When using the chat template (apply_chat_template) in lm-eval, performance degrades significantly. When running the evaluations without the chat template (using raw text generation formatting), the results perfectly align with the paper.

Furthermore, I was able to reproduce the paper results using both the vllm and hf backends. This confirms that the backend choice wasn't the problem, only the prompt formatting.

Here is a subset of my reproduction experiments (vLLM results):

Benchmark (Metric / Setting) Paper (1.4B R4) Ours (with chat template) Ours (NO chat template)
MMLU (acc, 5-shot) 67.35 66.54 67.46
BBH (exact_match (flexible), 3-shot CoT) 71.02 61.36 71.06
ARC-C (acc_norm, 25-shot) 60.92 57.34 60.41
GSM8K (exact_match (flexible), 3-shot CoT) 78.92 60.80 79.38

For reference, here are the environment versions I used for these runs:

  • vllm: 0.16.0
  • transformers: 4.57.6
  • lm-eval: 0.4.11

Sign up or log in to comment