dataset stringclasses 1
value | version stringclasses 1
value | generated_at stringclasses 1
value | stages dict |
|---|---|---|---|
CWI Public Corpus | 1.0.0 | 2026-10-02T22:48:33.116235+00:00 | {
"collected": 181,
"extract_failed": 0,
"extract_fetch_failed": 9,
"extracted": 0,
"filtered": 1,
"filter_reasons": {
"low_language_score": 1
},
"filter_passed": 171,
"dedup_removed": 12,
"dedup_kept": 159,
"dedup_ratio": 0.0702,
"dedup_threshold": 0.85,
"total_tokens": 31879,
"by_sourc... |
CWI Public Corpus
Version 1.0.0 — built by the CWI dataset-pipeline skill (FineWeb-shaped: trafilatura extraction → quality filters → MinHash dedup → BPE tokenization).
Contents
structured-record: 163 docstext-file: 9 docsunknown: 9 docs
Provenance
All documents come from public CWI sources only: the public
BlackLansky/cwi-catalog Hugging Face dataset, public GitHub Pages
(cumulativewebinc.github.io), and CWI's own published articles. Every row
carries source_url and collected_at. No private-repo or non-public data.
Pipeline
| stage | in | out |
|---|---|---|
| collect | — | 181 |
| extract (trafilatura) | 181 | 172 |
| quality filters | 172 | 171 |
| MinHash dedup (thr=0.85) | 171 | 159 |
| tokenize | 159 | 159 (31,879 tokens) |
Validation
See validation-report.json / validation-report.md in this repo.
Intended use
Pretraining / continued-pretraining and retrieval corpora for CWI music-catalog understanding, playlist-fit scoring, and trust-verdict classification. Not for production decisions without human review.
- Downloads last month
- 86