LLM Finetuning

Claude Code: /plugin marketplace add wshobson/agents → /plugin install llm-finetuning@claude-code-workflows

Eleven skills sequence the work: decide whether to fine-tune at all, build the eval harness, curate data, train, then gate the checkpoint. The routing skill opens with off-ramps — volatile facts belong in RAG, unsettled behaviour belongs in the prompt — because most requests that arrive as "fine-tune this" are cheaper to solve elsewhere. Evaluation comes first by design: the labelled traces that produce the golden set also produce training data, with every golden excluded by ID so leakage cannot quietly inflate the score. Promotion is a four-stage gate, and a checkpoint that wins on held-out goldens but loses the paired arena against the base model does not ship. Methods run from LoRA and QLoRA through DPO to GRPO with verifiable rewards, ending in GGUF or FP8 export.