SFT Desk - https://sft-desk.skillsafe.ai/

This app's agent prompt is a derivative work of one agent skill. It paraphrases and adapts its guidance;
it does not reproduce the skill or its reference files verbatim.

1. dataset-curation
   Source:  https://github.com/wshobson/agents
   Path:    plugins/llm-finetuning/skills/dataset-curation/SKILL.md (and references/formats-and-templates.md,
            references/synthetic-data.md)
   Author:  Seth Hobson (wshobson)
   Licence: MIT License - https://opensource.org/licenses/MIT
   Modifications: the skill's practice for fine-tuning datasets (the row shape by training method, the chat
   template applied before packing, loss masked to assistant turns, messages kept for the trainer rather than
   pre-rendered flat text, ShareGPT conversion, the 5-10 packed-sequence inspection, the 1,000-row SFT floor, the
   25% real-data collapse guard with replay rows counting as real, exact-then-semantic dedup, and the six-field
   dataset card with its exit checklist) was restated as deterministic browser checks (sftkit.js) and as the
   audit and card lanes of this app's prompt. The row-level checks (turn order, empty turns, conflicting answers,
   eval-split overlap, estimated lengths against max_length, refusals, preference/KTO/GRPO structure, personal
   data), the redacted sample, the clean export, the evidence and number checks against the page's own counts
   and the audit-to-card handoff are additions of this app.

The browser-side checks (sftkit.js), the reconciliation (recon.js) and the rest of the page are original to
this app. Model and library names are used only to describe formats; this app is not affiliated with any of
their publishers.