joshuasundance/mypo-4k-rfc

`mypo`数据集是一个专注于Python代码质量的DPO（Direct Preference Optimization）数据集。它包含三个主要列：`prompt`（原始提示）、`rejected`（存在linting错误的代码）和`chosen`（经过LLM修复后的代码）。数据集的目的是通过静态代码分析教导LLM在给定`prompt`时，选择`chosen`输出而非`rejected`输出，以提高代码生成质量。数据集来源于`iamtarun/python_code_instructions_18k_alpaca`，并通过`codellama/CodeLlama-7b-Python-hf`模型修复代码中的linting错误。

The `mypo` dataset is a preview version of a DPO dataset focused on Python code quality. It is derived from the `iamtarun/python_code_instructions_18k_alpaca` dataset and includes three main columns: `prompt` (from the original dataset), `rejected` (code from the original dataset with linting errors), and `chosen` (code rewritten by `codellama/CodeLlama-7b-Python-hf` to address linting errors). The dataset aims to train large language models (LLMs) to recognize and select the `chosen` output given the `prompt`, as it has better code quality than the `rejected` output. The creation process involves filtering out code with linting errors using static code analysis tools like `black`, `ruff`, and `mypy`, then rewriting these codes using LLM to correct the errors, and finally constructing the DPO dataset.