A dataset without a datasheet is a folder of CSV. ReadyPack takes your dataset plus the findings your audits already produced and hands back a publishable pack: a data-health score, a datasheet, Croissant metadata, a changelog, and a plain report.
You can run six honesty checks over a dataset and still ship something no one can reproduce, because the results live in six terminals and the dataset lives in a seventh. Packaging is the unglamorous step everyone skips. ReadyPack does that step and refuses to do it silently.
- Aggregates findings reports. Point it at the JSON envelopes from the other tools (EnrichAudit, LinguaCheck, SourceLock, WhyChanged, NuanceBench, FilterGap, DecayWatch) and it folds them into one report, tagging each finding with the audit it came from.
- Scores data health. A single 0–100 score with a letter grade. Penalties
are log-scaled (high=6, medium=3, low=1), so the first problems cost the
most and a long tail of minor notes can't sink an otherwise good dataset.
summary.health_score_detailexposes the per-source components, weights, and a confidence rating; the report includes a FAIR self-assessment. - Writes the datasheet. Composition, fields, per-audit findings, and a limitations section built from the warnings your audits emitted, as both Markdown and a machine-readable Datasheets-for-Datasets JSON document.
- Emits Croissant 1.1 metadata.
dct:conformsTo, a version, a generation timestamp, inferred field data types, a SHA-256 ofcleaned.csv, and a PROV-O-style record of which audit tool, version, and status contributed each finding. A schema.org/Dataset JSON-LD document is included too. - Keeps a changelog. Every reviewed removal and the packaging run are timestamped events.
- Packages it up.
cleaned.csv, the datasheets,croissant.json,dataset.jsonld,findings.json,changelog.json,run.json, andreport.md, zipped intopackage.zip. - Never silently deletes. Rows only leave the export when you hand over ids you already reviewed. The tool's default is to change nothing.
It is not a checker. It does not look for leaks, gaps, PII, or bots — the other tools in the family do that. ReadyPack's job starts once you have findings and ends with a documented, packaged dataset.
Python 3.10+, no runtime dependencies.
git clone https://github.com/hexsyro/ReadyPack.git
cd ReadyPack
python -m pip install .For development:
python -m pip install -e ".[dev]"
python -m pytestreadypack dataset.csv \
--findings filtergap-report.json \
--findings linguacheck-report.json \
--name "product-mentions-2026" \
--description "Mentions collected for the launch study." \
--drop-id r003 \
--out ./packWithout --findings, ReadyPack still documents and packages the dataset — it
just has no audits to fold in. Without --out, it prints the report and writes
nothing.
| Flag | What it does |
|---|---|
--findings FILE |
A findings report (JSON envelope) from another tool; repeatable |
--out DIR |
Write the package (cleaned.csv + docs + package.zip) to DIR |
--name NAME |
Dataset name for the datasheet and Croissant metadata |
--description TEXT |
Dataset description |
--drop-id ID |
Drop a reviewed row id from the packaged export; repeatable |
--config FILE |
JSON config; explicit flags win |
--format {table,json,csv,md} |
Output format (default table) |
--output FILE |
Write the report to FILE instead of stdout |
--fail-on {none,low,medium,high} |
Exit 1 if any finding is at or above this severity |
--version |
Print version and exit |
{
"tool": "readypack",
"version": "0.2.0",
"status": "success",
"findings": [
{
"severity": "high",
"field": "bot_confidence",
"message": "Mention scored 0.95 bot/spam confidence (threshold 0.50).",
"record_id": "r003",
"metadata": {"check": "bot_confidence", "confidence": 0.95, "audit": "filtergap"}
}
],
"summary": {
"result_schema_version": "1.1",
"skipped": false,
"skip_reason": null,
"checked": 5,
"name": "product-mentions-2026",
"dataset": {"fields": ["id", "text", "language", "split"], "row_count": 4, "original_row_count": 5},
"sources": {"filtergap": {"findings": 1, "by_severity": {"high": 1, "medium": 0, "low": 0}, "score": 78.9, "grade": "B"}},
"findings_total": 1,
"health_score": {"score": 78.9, "grade": "B", "weights": {"high": 6.0, "medium": 3.0, "low": 1.0}},
"health_score_detail": {"components": {}, "weights": {"high": 6, "medium": 3, "low": 1}, "confidence": "high"},
"drop_count": 1,
"unknown_drop_ids": [],
"datasheet": "# Datasheet: product-mentions-2026\n...",
"datasheet_json": "{\"Motivation\": ...}",
"dataset_jsonld": "{\"@type\": \"Dataset\", ...}",
"findings_json": "[{\"severity\": \"high\", \"tool\": \"filtergap\", ...}]",
"croissant": {"@type": "sc:Dataset", "name": "product-mentions-2026"},
"report_md": "# ReadyPack report: product-mentions-2026\n...",
"changelog": [{"at": "2026-10-19T00:00:00+00:00", "actor": "readypack", "event": "Packaged 4 row(s)..."}]
},
"warnings": ["Review every finding before publishing — ReadyPack never silently deletes rows."]
}The pack on disk (--out) contains cleaned.csv, datasheet.md,
datasheet.json, croissant.json, dataset.jsonld, findings.json,
changelog.json, run.json, report.md, and package.zip. The
"package written to …" notice goes to stderr so --format json stays valid.
readypack dataset.csv --findings audit.json --fail-on high || exit 1Exit codes: 0 ran fine, 1 gate tripped, 2 input broken.
from readypack import run, health_score
result = run(rows, {"name": "demo", "reports": [filtergap_report]})
result.summary["health_score"]["score"]
health_score([{"severity": "high"}]) # (78.9, "B")- It trusts your reports. ReadyPack folds in the findings it is given. If an audit was wrong, the score is wrong. It scores findings, not truth.
- The score is a summary, not a judgment. A 92 with one high finding is still a dataset with a high finding. Read the findings, not just the grade.
- Dropping rows is manual on purpose. You pass the ids. ReadyPack will not infer them, because the whole point is that a human decided.
- Croissant output is focused, not exhaustive. It covers the core dataset, record-set, distribution, provenance, and checksum concepts; it is not a complete implementation of every Croissant 1.1 extension.
- The license is a placeholder. Replace
licenseindataset.jsonldandcroissant.jsonbefore publishing.
- Full Croissant 1.1-style metadata:
dct:conformsTo, dataset version,generatedAt, inferredrecordSet[].fielddata types,distribution[].sha256, and a PROV-O-stylewasGeneratedBy/wasAttributedToblock per source report. - Machine-readable
datasheet.json,dataset.jsonld(schema.org/Dataset), andfindings.json, written to disk, zipped intopackage.zip, and available as text in the result summary.write_package(..., extra_files=...)adds caller documents such asrun.json(nested relative paths allowed,..rejected). summary.health_score_detailwith per-tool components, weights, and confidence, plus a FAIR self-assessment table in the Markdown report.- Result-contract fields
result_schema_version="1.1",skipped, andskip_reason;RESULT_SCHEMA_VERSIONandToolResultexported from the package. - Changelog events for each reviewed drop id in addition to the packaging event.
--format jsonstdout is no longer corrupted by the "package written to …" notice; the notice goes to stderr.- Uppercase finding severities (e.g.
High) now receive the correct health penalty. - Duplicate audit tool names no longer overwrite each other; they get a numeric suffix and a warning.
- Field lists and the CSV header use the union of keys across all rows, not just row 1.
- Unknown drop ids produce a warning (first 10 shown) and
summary["unknown_drop_ids"]. - Invalid config value types are rejected by the CLI with exit 2 and a clear stderr message; library calls coerce defensively instead of raising.
- Pylint CI now runs on Python 3.10–3.13.
- Croissant 1.1 + PROV-O provenance — https://mlcommons.org/2026/02/croissant-1-1-standard
- Croissant spec (version + checksums) — https://docs.mlcommons.org/croissant/docs/croissant-spec-1.0.html
- Datasheets for Datasets — https://www.microsoft.com/en-us/research/wp-content/uploads/2019/01/1803.09010.pdf
- Croissant-RAI — https://proceedings.neurips.cc/paper_files/paper/2024/file/9547b09b722f2948ff3ddb5d86002bc0-Paper-Datasets_and_Benchmarks_Track.pdf
- schema.org/Dataset JSON-LD — https://developers.google.com/search/docs/appearance/structured-data/dataset
- HN "which code ran" — https://news.ycombinator.com/item?id=49510632
MIT. See LICENSE.