Dataset Curation

Before you touch a training run, this skill makes you do the boring part properly: format rows for the method you picked, apply the chat template before packing, mask loss to assistant turns, and decode a handful of packed sequences by hand, since packing bugs don't crash anything and only show up later in eval quality. It also insists on at least 25% real data in the final mix to guard against synthetic collapse, and it won't hand off to a run until a six-field dataset card is filled in. The honest take is that the ~1,000-row SFT floor and the 25% real-data rule are sensible defaults rather than hard science, and the skill is most useful as a checklist you'd otherwise skip under deadline pressure.

skills.sh installs: 3K
GitHub stars: 42.4K
last updated: Jul 22, 2026

View on GitHub

Install

$ npx -y skills add wshobson/agents --skill dataset-curation --agent claude-code

Installs into .claude/skills of the current project.

Files

Dataset Curation

This skill assumes finetuning-method-selection already routed here — the next step is preparing data, not choosing a method. What follows: format selection by target method, the template/packing mechanics behind the most common silent training failures, rules for mixing in synthetic data without collapse, and the dataset card that closes out Phase 2 before a run starts.

Input: raw examples (demonstrations, preference judgments, or task prompts) plus a routing decision from finetuning-method-selection. Output format: a formatted, packed, validated JSONL dataset plus a completed dataset card — the Phase 2 artifact /finetune checks before launching training.