# dropoutt

> Pre-flight checks for LLM training data.

dropoutt is a local command-line tool. Point it at a folder of training data. `dropoutt scan` reports what would break a training run (empty loss masks, truncation, duplicates, benchmark contamination, language damage, personal data, token cost) and writes one report as HTML, Markdown and JSON. `dropoutt atlas` places the same folder on atlas-v3, a fixed map of public training data, and reports where it sits, where it piles up and what it never reaches.

Version 1.4.0. Python 3.10 or newer. Apache-2.0. One install, no extras, no model required. Everything runs on your machine.

```
pip install dropoutt
dropoutt scan ./data
dropoutt atlas --model atlas-v3 ./data
```

Pass `--model` to a scan to unlock token and loss-mask checks. Pass `--target sft` to turn blocking findings into exit code 10. Without `--target`, a scan reports and exits 0. `dropoutt atlas` never fails a build.

## Docs

- [Get started](https://dropoutt.com/docs): install, first scan, what you get with no flags, the map
- [Command reference](https://dropoutt.com/docs/commands): scan and atlas flags, fetch, models, benchmarks, exit codes, CI, dropoutt.toml, environment variables
- [Reading the report](https://dropoutt.com/docs/reports): report.html / report.md / report.json / findings.jsonl / fingerprint.json, and the atlas.* files
- [Atlas](https://dropoutt.com/docs/atlas): atlas-v3, 4,096 hand-named cells in 256 subject areas, fitted once on 163 million public records; atlas-v2 and atlas-v2-lite
- [Check catalog](https://dropoutt.com/docs/checks): 36 checks, ids never renumbered

## Source

- [GitHub](https://github.com/ckproduces/dropoutt-cli)
- [PyPI](https://pypi.org/project/dropoutt/)
- [Changelog for 1.4.0](https://github.com/ckproduces/dropoutt-cli/blob/v1.4.0/CHANGELOG.md)

## Optional

- [llms-full.txt](https://dropoutt.com/llms-full.txt): the same, plus the full check catalog and the map's products