minimax-m27-cc

# MiniMax-M2.7 Claude Code Traces Clean Claude Code proxy traces collected from `MiniMaxAI/MiniMax-M2.7` for DFlash draft-model training. ## Contents Rows are pre-rendered with the MiniMax-M2.7 chat template and are intended to be loaded by `prepare_minimax.py`. Important columns: - `prompt_text`: MiniMax chat-template prompt ending at the assistant generation prompt. - `target_text`: assistant continuation to supervise. - `trace_run_id`, `trace_task_id`, `request_id`: provenance. - `request_body_bytes`, `target_char_count`, `has_tool_call`: filtering/debug metadata. ## Filtering Source run prefix: `minimax-m27-20260510-054353-bypass` Included: - `/v1/messages?beta=true` rows only. - HTTP 200 rows only. - Rows whose final Claude Code task status was `ok`. - Rows with valid tool-call JSON. - Rows with non-empty rendered target text. Excluded: - count-token rows. - transport-miss rows. - rows from max-turn tasks. - malformed or empty responses. Final row count: `15144`. ## Usage ```bash python prepare_minimax.py \ --input z-lab/minimax-m27-cc \ --output /path/to/tokenized \ --model MiniMaxAI/MiniMax-M2.7 \ --max_length 80960 \ --num_workers 32 ``` Some late-turn rows are longer than 80k tokens. `prepare_minimax.py` will skip rows where truncation leaves no supervised target tokens. For preserving the longest rows, use a larger training max length.