README.md

metadata

title: Wechat Style Sft
emoji: 🌍
colorFrom: yellow
colorTo: indigo
sdk: gradio
pinned: false
license: apache-2.0
short_description: QLoRA fine-tuning demo on WeChat essays for style adaptation
sdk_version: 5.46.1

WeChat Style SFT (QLoRA) — Minimal Space

Steps:

1. Upload your two JSONLs into data/:
   • data/sft_train.jsonl
   • data/sft_val.jsonl
2. Start the Space with CPU Basic (free) to verify requirements install.
3. Switch to ZeroGPU / T4 in Settings → Hardware.
4. Tab "Train" → Start Training (default 500 steps, Qwen2.5-0.5B).
5. After training, a lora_adapter/ dir is saved.
6. Switch back to CPU Basic; Tab "Inference" → Load → Generate.

Privacy/Compliance:

• Use your own copyrighted text only.
• Do not upload full articles to public repos.
• Keep this Space Private under NUS org.