We Got Claude to Fine-Tune an Open Source LLM

Abstraction: Hugging Face skill enables Claude Code to submit and manage LLM training jobs

Key points:

• hf-llm-trainer skill lets Claude Code validate datasets, select GPU hardware, submit training jobs to Hugging Face Jobs, monitor via Trackio, and push finished models to Hub
• Supports SFT, DPO, and GRPO (RL with verifiable rewards); models from 0.5B to 70B; LoRA auto-applied for models above 3B
• Hardware cost examples: t4-small for <1B (~$0.30 for 20 min), a10g-large with LoRA for 3-7B (~$15-40 production)
• Compatible with Claude Code, OpenAI Codex, and Gemini CLI; installed via /plugin install hf-llm-trainer@huggingface-skills
• Post-training GGUF conversion supported via a separate job that merges LoRA adapters and applies quantization
• Dataset validation runs on CPU (fractions of a penny) before committing to GPU runs; DPO requires exact chosen/rejected column names

Connections: Hugging Face · Claude Code · Fine Tuning · Reinforcement Learning · Agentic Coding

Source: https://huggingface.co/blog/hf-skills-training