Fine-Tuning Method Selection
Decides whether fine-tuning is worth it, then routes to the right method and model size class.
- Repository
- wshobson/agents
- Stars
- 42,095
- Licence
- MIT
- Files
- 3 files
What it does
This skill makes the agent act as a router for fine-tuning projects. Before any training starts, it checks whether retrieval (RAG) or prompt engineering would solve the problem more cheaply, then picks a method such as SFT, DPO/KTO, or GRPO based on the kind of data available. It also gives rough sizing guidance on model size and memory needs.
Great for
- Teams starting a fine-tuning project who are not yet sure whether they need training at all
- Engineers choosing between preference optimization (DPO-family) and reinforcement methods (GRPO/RLVR)
- Anyone who wants a memory and model-size estimate before committing to a training run
Skip it if
- Someone who has already chosen a method and needs step-by-step training code; the skill points to other skills for that
- Facts that change often (prices, news, docs): the skill itself says to use RAG instead, since a fine-tuned model goes stale quickly
- Teams without an evaluation setup: the skill says an eval harness must exist before any training, so it is not a shortcut around that work
Try asking
- “I have 2,000 support transcripts showing ideal agent replies. Should I fine-tune, and with which method?”
- “Our pricing page changes weekly. Do we need fine-tuning or RAG?”
- “We have thumbs-up and thumbs-down ratings on model answers but no paired comparisons. Which method fits?”
- “Roughly how much GPU memory would an 8B model need for LoRA fine-tuning?”
Install in 3 steps
- 1
Download the skill
The ZIP contains the 3 original files, the MIT licence and an attribution note.
finetuning-method-selection.zip - 2
Add it to your agent
Claude Code (all projects)
mkdir -p ~/.claude/skills && unzip finetuning-method-selection.zip -d ~/.claude/skills/Claude Code / Agent SDK (this project only)
mkdir -p .claude/skills && unzip finetuning-method-selection.zip -d .claude/skills/Claude apps (web/desktop): Settings → Capabilities → Skills → Upload skill, and choose the ZIP.
- 3
Ask for the task
No special command needed — the agent reads the skill's description and loads it when your request matches. Try one of the prompts above.
Get better results
- Describe your data shape (demonstrations, preference pairs, thumbs up/down, or pass/fail checks) and the volume of domain text; the routing depends on both.
- Check the 'last verified' date in the model catalog before accepting any model recommendation, since model rankings change quickly.
- Expect the skill to point you to companion skills (such as lora-qlora-recipes or grpo-rlvr-training) for the actual implementation details.
What's inside
- references/memory-math.md
- references/model-catalog.md
- SKILL.md
Redistributed unmodified from wshobson/agents under the MIT licence. The assessment was drafted with AI from the skill's files on 2026-10-29. Skills can instruct your agent to run code — read SKILL.md and any scripts before installing.