Memorization or Generation of Big Code Model Leaderboard

Context

The growing number of code models released by the community necessitates a comprehensive evaluation to reliably benchmark their capabilities. Similar to the 🤗 Open LLM Leaderboard, we selected two common benchmarks for evaluating Code LLMs on multiple programming languages:

HumanEval - benchmark for measuring functional correctness for synthesizing programs from docstrings. It consists of 164 Python programming problems.
MultiPL-E - Translation of HumanEval to 18 programming languages.
Throughput Measurement - In addition to these benchmarks, we also measure model throughput on a batch size of 1 and 50 to compare their inference speed.

Benchmark & Prompts

HumanEval-Python reports the pass@1 on HumanEval; the rest is from MultiPL-E benchmark.
For all languages, we use the original benchamrk prompts for all models except HumanEval-Python, where we separate base from instruction models. We use the original code completion prompts for HumanEval for all base models, but for Instruction models, we use the Instruction version of HumanEval in HumanEvalSynthesize delimited by the tokens/text recommended by the authors of each model (we also use a max generation length of 2048 instead of 512).

Figure below shows the example of OctoCoder vs Base HumanEval prompt, you can find the other prompts here.

- An exception to this is the Phind models. They seem to follow to base prompts better than the instruction versions. Therefore, following the authors' recommendation we use base HumanEval prompts without stripping them of the last newline. - Also note that for WizardCoder-Python-34B-V1.0 & WizardCoder-Python-13B-V1.0 (CodeLLaMa based), we use the HumanEval-Python instruction prompt that the original authors used with their postprocessing (instead of HumanEvalSynthesize), code is available [here](https://github.com/bigcode-project/bigcode-evaluation-harness/pull/133).

Evalution Parameters

All models were evaluated with the bigcode-evaluation-harness with top-p=0.95, temperature=0.2, max_length_generation 512, and n_samples=50.

Throughput and Memory Usage

Throughputs and peak memory usage are measured using Optimum-Benchmark which powers Open LLM-Perf Leaderboard. (0 throughput corresponds to OOM).

Scoring and Rankings

Average score is the average pass@1 over all languages. For Win Rate, we find model rank for each language and compute num_models - (rank -1), then average this result over all languages.

Miscellaneous

#Languages column represents the number of programming languages included during the pretraining. UNK means the number of languages is unknown.

How to submit models/results to the leaderboard?

We welcome the community to submit evaluation results of new models. These results will be added as non-verified, the authors are however required to upload their generations in case other members want to check.

1 - Running Evaluation

We wrote a detailed guide for running the evaluation on your model. You can find the it in bigcode-evaluation-harness/leaderboard. This will generate a json file summarizing the results, in addition to the raw generations and metric files.

2- Submitting Results 🚀

To submit your results create a Pull Request in the community tab to add them under the folder community_results in this repository:

Create a folder called ORG_MODELNAME_USERNAME for example bigcode_starcoder_loubnabnl
Put your json file with grouped scores from the guide, in addition generations folder and metrics folder in it.

The title of the PR should be [Community Submission] Model: org/model, Username: your_username, replace org and model with those corresponding to the model you evaluated.

⭐ Memorization or Generation of Big Code Models Leaderboard