3. OpenAI Clean (LLM Cleanup)

Goal: clean the distilled SFT using an OpenAI-compatible Chat Completions API, to reduce tool/technical/search traces and produce a more stable training dataset.

---

Configure an OpenAI-compatible endpoint

Edit setting.jsonc:

• data_args.clean_set_args.openai_api.api_base
• data_args.clean_set_args.openai_api.api_key
• data_args.clean_set_args.openai_api.model_name
• data_args.clean_set_args.openai_api.clean_workers

Note: api_base should point to .../v1/chat/completions.

---

Run (defaults to latest distill output)

python cli.py data openai-clean

Outputs:

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

---

Small-scale validation (recommended)

python cli.py data openai-clean --max-samples 200

---

System prompt injection (optional)

By default, it reads data_args.openai_sft_system_prompt.

• Disable: --no-base-prompt
• Override: --base-prompt / --base-prompt-file