Reading the report
The terminal, HTML, Markdown, and JSON outputs present the same scan results in the same order. The map writes its own set of files.
Dataset composition
The composition section summarizes language distribution, detected training layouts, embedded chat templates, average record size, dataset licences, and which datasets repeat each other.
Findings
Each finding includes the affected count, denominator, estimated wasted tokens when available, recommended action, and evidence from the dataset. Pass --target to enable blocking findings and exit code 10.
Checks that could not run are listed at the end with the one flag that would unlock each.
Token estimates
With --model, dropoutt reports exact token counts for that tokenizer. Without it, dropoutt estimates counts for 5 common tokenizers from a sample and reports the sampling interval.
Token estimates are calculated per dataset and then combined, preserving differences between languages and data sources.
Atlas coverage
Coverage is reported by dropoutt atlas. The command writes atlas.html, atlas.md, and atlas.json in the same output directory, with these sections:
| Section | What you learn |
|---|---|
| The map | Every subject area you reached and its subregions, each with its density against the map. |
| What the map says | Sentences that clear both a size gate and a significance gate. |
| What you have most of | The crowded places, each described by your own record nearest its centre. |
| What you have least of | The sparsest places you reach. Reaching a place is not covering it. |
| Farthest from the map | Where your mix differs most from the map, in either direction. |
| Off the map | Records unlike the reference geography, with a diagnosis. Short records are the usual cause before subject is. |
Output files
| File | For | What it holds |
|---|---|---|
| report.html | review | Self-contained report that opens directly from disk. No CDN, no web fonts. |
| report.md | PRs and tickets | Scan results as markdown. |
| report.json | automation | Scan results as one JSON document. |
| findings.jsonl | scripts | One JSON object per finding, with counts, rates and evidence. |
| fingerprint.json | comparison | Comparable dataset facets without record excerpts. |
| atlas.html / .md / .json | coverage | Coverage results written by dropoutt atlas. The JSON follows schema dropoutt.atlas/1. |
--no-evidence before sharing reports outside the dataset trust boundary. It is applied once, where the content is assembled, so every format honours it.