tags:
- code
- evaluation
- code llm
size_categories:
- n<1K
Abstract
Driven by the surge in code generation using large language models (LLMs), numerous benchmarks have emerged to evaluate these LLMs' capabilities. We conducted a large-scale human evaluation of HumanEval and MBPP, two popular benchmarks for Python code generation, analyzing their diversity and difficulty. Our findings unveil a critical bias towards a limited set of programming concepts, neglecting most of the other concepts entirely. Furthermore, we uncover a worrying prevalence of easy tasks that can inflate model performance estimations. To address these limitations, we propose a novel benchmark, PythonSaga, featuring 185 hand-crafted prompts in a balanced representation of 38 programming concepts across diverse difficulty levels. The robustness of our benchmark is demonstrated by the poor performance of existing Code-LLMs. The code and dataset are openly available to the NLP community at https://github.com/PythonSaga/PythonSaga.
PythonSaga
This dataset follows the rules and diversity of template suggested in the paper "PythonSaga: Redefining the Benchmark to Evaluate Code Generating LLM" The goal is to make benchmarks better at assessing Code Generating Language Models (LLMs).| Model | Size | Pass@1 | Pass@10 |
|---|---|---|---|
| StarCoderBase | 7B | 0.0029 | 0.0149 |
| StarCoder2 | 7B | 0.0024 | 0.0217 |
| Code Llama | 7B | 0.0067 | 0.0472 |
| CodeQwen1.5-Chat | 7B | 0.0059 | 0.0497 |
| Nxcode-CQ-orpo | 7B | 0.0058 | 0.0523 |
| Mistral-Instruct-v0.1 | 7B | 0.0140 | 0.0552 |
| Code Llama Instruct | 7B | 0.0178 | 0.0744 |
| Deepseek Coder Instruct | 6.7B | 0.0137 | 0.0889 |
| Code Llama Python | 7B | 0.0240 | 0.0979 |
| Llama 3 | 8B | 0.0370 | 0.1125 |
| Phi-2 | 2.7B | 0.0302 | 0.1187 |
| OpenCodeInterpreter-DS | 6.7B | 0.0259 | 0.1206 |
| Deepseek Coder | 6.7B | 0.0343 | 0.1415 |
| Code Llama Python | 13B | 0.0405 | 0.1514 |
| GPT-3.5 | NA | 0.0724 | 0.2384 |
| GPT-4 | NA | 0.1243 | 0.3311 |
Comparison between open and closed-source models on PythonSaga. We use the number of samples (n) as 20 for both open and closed-source models.