SFT Desk - check a fine-tuning dataset before the run, then write its dataset card

Paste or drop the JSONL for one SFT, DPO, ORPO, KTO or GRPO run: your browser checks every row for free (shape against the method, turns and roles, empty answers, flat text and pre-rendered templates that defeat assistant-only loss, ShareGPT rows, duplicates, eval-split overlap, max_length, the 1,000-row SFT floor, the 25% real-data floor, refusals, personal data) and builds a clean export; then an audit judges what a redacted sample of rows teaches and names the rows to drop, and the card lane writes the six-field dataset card the run is gated on. Derived from @wshobson/dataset-curation (Seth Hobson, wshobson/agents, MIT).

Every public app is built from a security-scanned skill and must pass a clean scan — skill and frontend — before it can be listed. Have a skill of your own? Turn it into an app — or read the step-by-step walkthrough.