Keep4o

Keep4o is a pipeline for “GPT-4o style alignment”: take your own ChatGPT export, distill it into trainable SFT (text-first), optionally run an LLM-based cleanup via an OpenAI-compatible endpoint, then fine-tune/distill.

This route focuses on output structure, clarification behavior, refusal habits, tone, and tool-call tendencies. It is not the “digital avatar” route. If you want to fine-tune personal style from QQ/TG data, follow the Quick Start.

---

Outputs

• runs/openai-distill/<run_id>/sft/text.jsonl
• runs/openai-clean/<run_id>/sft/train.jsonl

---

Privacy

• ChatGPT exports are highly private; keep them under data/openai-export/
• Do not commit data/ or runs/ to git

---

Next

Start from “Export ChatGPT Data” in the sidebar.