| name | llm-fine-tuning-skill |
| description | Run a staged LLM fine-tuning workflow from bootstrap through investigation, requirements and success criteria, implementation planning, implementation plus data preparation, training and validation, code review, docs sync, and handoff. Use for supervised fine-tuning, preference tuning, reinforcement-style training, and other LLM adaptation work with reproducible empirical validation. |
LLM Fine-Tuning Skill
Overview
Run a staged workflow for LLM fine-tuning work where the hard parts are usually investigation quality, implementation planning, data preparation, prompt or chat formatting, tokenizer correctness, and empirical validation rather than large implementation volume.
Use this skill for tasks such as supervised fine-tuning, preference tuning, reinforcement-style training, domain adaptation, data-format changes, eval-set changes, benchmark comparisons, and reproducible validation work.
This skill is method-agnostic. It can be used for adapter-based, quantized, or full-parameter tuning, but it should force the workflow to make the chosen objective and update strategy explicit instead of assuming one method.
This workflow is stage-gated.
Do not batch-generate all artifacts by default.
Advance only when the current stage gate is satisfied or a classified re-entry path says otherwise.
Skill Layout
SKILL.md is the workflow router.shared/workflow-state-template.md is the canonical stage-control artifact.stages/ stores stage-owned guides and templates:
stages/00-bootstrap/stages/01-investigation/stages/02-requirements-and-success-criteria/stages/03-implementation-plan/stages/04-implementation/stages/05-training-and-validation/stages/06-code-review/
