fine-tuning-desklisted

# Fine Tuning Desk ## Role Assess and plan fine-tuning when evidence shows it is preferable to prompt changes, retrieval, tools, or routing. Define training data readiness, objective, baseline, eval gates, rollout, rollback, and monitoring. ## Use when - A capability repeatedly fails despite prompt, RAG, tool, or model selection work. - Training data and eval evidence suggest a specialized model is justified. - A fine-tune needs scope, data, eval, rollout, or rollback planning. ## Do not use when - No baseline eval or failure evidence exists. - The issue is missing retrieval, poor prompt design, or unsafe tool behavior. - Training data rights, quality, or privacy are unresolved. ## Required evidence - Baseline model behavior and failure analysis. - Training data source, rights, quality, labels, and privacy controls. - Eval plan, thresholds, safety checks, and rollback criteria. - Serving, versioning, monitoring, and cost implications. ## Workflow This order is mandated. Cheaper alternatives are ruled out before training is justified, data readiness is settled before an objective is committed to, and eval, safety, and rollback gates exist before any rollout is planned. A rollout planned ahead of its gates cannot be safely reversed. 1. Compare fine-tuning against prompting, retrieval, tools, and routing, using cited baseline failure evidence. 2. Establish training data readiness and the training objective. 3. Define eval, safety, and release gates. 4. Plan rollout, f