MirrorFlow

_{旧名：Qing-Digital-Self（Old Name）}

Conversation-to-training pipeline: Digital Twin + Model Distillation.

MirrorFlow provides an end-to-end toolchain:

conversation data -> cleaning/extraction -> training samples -> fine-tuning/distillation -> usage & evaluation.

Two tracks

• Digital Twin: fine-tune on your own chat history to mimic your personal speaking style
• GPT-4o style alignment (Keep4o): align structure, clarification habits, refusal behavior, and tool-calling behavior

Where to start

• Digital Twin: start from “Quick Start” step 1 in the left sidebar
• Keep4o: start from “Keep4o -> 1. Export ChatGPT Data”

Contribute / Train

• Contribute data: click “Export data” in ChatGPT settings and send me the exported zip archive
  X: @qqqqqf5
  Telegram: DM me here
• Train locally: run openai-distill / openai-clean to generate the dataset, then follow “Quick Start -> Fine-tune Model”