sft-trainer

---

title: Qwen3-Coder-Next Uncensored Fine-Tuning emoji: "\U0001F525" colorFrom: red colorTo: yellow sdk: docker app_port: 7860 suggested_hardware: a100-large suggested_storage: large pinned: true

license: apache-2.0

Qwen3-Coder-Next Uncensored Fine-Tuner

Fine-tune Qwen/Qwen3-Coder-Next (80B MoE / 3B active) using QLoRA to remove refusal behavior.

Setup

1. Create a HF Space with Docker SDK and a100-large hardware
2. Add Secrets in Space Settings:
   • HF_TOKEN — your Hugging Face write token
   • WANDB_API_KEY — (optional) for training metrics logging
3. Upload this repo to the Space
4. The Gradio UI will let you configure and launch training

Features

• QLoRA 4-bit fine-tuning (fits on single A100 80GB)
• Multiple uncensored dataset options built-in
• Custom dataset upload support
• Real-time training metrics in UI
• One-click LoRA merge + push to Hub
• Abliteration support (no training needed)

Hardware Requirements

Method	VRAM	Recommended HF Hardware
QLoRA 4-bit	~45-60GB	A100 Large (80GB)
Abliteration	~45-60GB	A100 Large (80GB)