Pass@1 score by oracle code is only 36.36%

[sfc-gh-hhan]

Despite following this, the environmental setting is still problematic. I ran the Pass@1 evaluation using the oracle code snippets from data.jsonl:

Running pass@1 for local_infilling oracle_greedy
TODO Completions:  275
100%|██████████| 275/275 [13:33<00:00,  2.96s/it]
pass_at_1: 36.36363636363637%

My conjecture is that, as some repositories do not specify their package versions, the environmental set up is prune to be stale.

A best way is providing a docker like bigcode-evaluation-harness. I notice that you are making a docker. I wonder when would it be available.

Before that, you may share the specific pip version and update the repo's requirements.txt with specific package versions by pip freeze.

[pldlgb]

Does this mean that the results of the paper are also not credible? Because a large part cannot be passed, the model will fail even if it is done correctly, resulting in unfair comparisons.

[sfc-gh-hhan]

As the GPT-4 Pass@1 score gets higher than the reported one (20.73 to 26.54) after fixing some environmental issues, I guess the environmental setup for the reported result in the paper is also problematic.

[LJ2lijia]

Thank you for your suggestions. We would like to clarify three points:

(1) First, in our experimental environment, we ensured that all reference code could pass all test cases. This result demonstrates that our execution environment provides the required packages to support the reference code. We believe our experimental results are credible.

(2) Recently, we found that the code generated by GPT-4 or other models may rely on additional third-party libraries. These additional libraries are not in the requirement.txt and can only be configured manually. We will address this problem in future work.

(3) Your suggestion is great @sfc-gh-hhan. We will release dock files to facilitate the evaluation. Thank you.

[sfc-gh-hhan]

Thank you for your response:

(1, 3) It's a relief that all the reference code passed all the test cases. I am really looking forward to running my experiments on your benchmark. May I ask for a rough timeline regarding the Docker release?

(2) That is an interesting point. Should we regard them as incorrect despite not providing the explicit list of installed packages? I'm happy to hear your thoughts.

[iwangjian]

Meet the same issue that reference code cannot pass all test cases (Pass@1 is only about 36%). May I know when the docker files will be released? Or, could you please share the specific pip version and update the repo's requirements.txt?

[royzhao]

Thank you for your suggestions. We would like to clarify three points:

(1) First, in our experimental environment, we ensured that all reference code could pass all test cases. This result demonstrates that our execution environment provides the required packages to support the reference code. We believe our experimental results are credible.

(2) Recently, we found that the code generated by GPT-4 or other models may rely on additional third-party libraries. These additional libraries are not in the requirement.txt and can only be configured manually. We will address this problem in future work.

(3) Your suggestion is great @sfc-gh-hhan. We will release dock files to facilitate the evaluation. Thank you.

Excuse me, may I ask when the Dockerfile will be released? I'm currently having trouble reproducing the experimental results, and I'm wondering if it might be due to environmental factors.