claude_converter: Convert Claude Code Sessions into Fine-tuning Data for Local Models
English summary
A developer released claude_converter, an open-source tool that converts Claude Code session .jsonl files into the messages format accepted by fine-tuning frameworks like TRL/SFTTrainer, Axolotl, and LLaMA-Factory (ShareGPT format). It includes a clean_messages() helper to strip tool-use blocks and an inspect_session() function for token counts and breakdowns. The tool has zero dependencies and can be installed via `uv pip install claude-converter`. Users are advised to filter sessions to only those where the final assistant turn solved the problem before training.
Chinese summary
开发者发布了 claude_converter，一款开源工具，可将 Claude Code 会话的 .jsonl 文件转换为 TRL/SFTTrainer、Axolotl 和 LLaMA-Factory（ShareGPT 格式）等框架可接受的 messages 格式。该工具提供 clean_messages() 辅助函数以去除工具使用块，以及 inspect_session() 函数用于显示 token 计数和结构分解。工具无外部依赖，可通过 `uv pip install claude-converter` 安装。建议用户仅筛选出最终助手回复确实解决了问题的会话再进行训练。
Key points
Converts Claude Code session .jsonl files (from ~/.claude/projects/) into a standard messages format for fine-tuning.
将 Claude Code 会话 .jsonl 文件（位于 ~/.claude/projects/）转换为标准的微调用 messages 格式。
Outputs compatible with TRL/SFTTrainer, Axolotl, and LLaMA-Factory (ShareGPT format).
输出兼容 TRL/SFTTrainer、Axolotl 和 LLaMA-Factory（ShareGPT 格式）。
Includes clean_messages() to remove tool-use blocks, and inspect_session() for token count and session structure overview.
包含 clean_messages() 去除工具使用块，以及 inspect_session() 查看 token 计数和会话结构概览。
Zero external dependencies; install via `uv pip install claude-converter`.
零外部依赖；通过 `uv pip install claude-converter` 安装。
Authors caution users to filter sessions, keeping only those where the final assistant turn successfully solved the task.
作者提醒用户需筛选会话，仅保留最终助手回复成功解决了任务的会话。
Repository: https://github.com/FredyRivera-dev/claude_converter
仓库地址：https://github.com/FredyRivera-dev/claude_converter