storagestuffnobody/finepdf-subsampled-cleaned

FinePDF cleaned copy

Status: complete. Published source shards: 1007/1007. This is an independently saved derivative; the source and its embeddings remain unchanged.

Source: storagestuffnobody/finepdf-subsampled. The bucket audit documents the earlier 64-bucket review and full-corpus repetition measurements.

Removal uses narrow validated ebook-spam/report-sales templates, individually reviewed ads, and the score>=0.95 cores of fine embedding buckets 539 and 1539. A fresh sample from those cores contained 126 advertisements among 128 reviewed documents. All reviewed useful exceptions are protected. This is a sampled, high-purity cut, not a guarantee that every unseen removed member is spam. Mixed boundaries and other broad buckets are retained.

Repetition repair is restricted to adjacent exact document-scale loops, individually supported block repairs, and six reviewed caption/placeholder loops. All distinct whitespace-normalized paragraphs are retained. Broad deduplication of recurring legal clauses, liturgical responses, or other contextual text is not applied. Exact character spans, retained-copy witnesses, hashes and original metadata are in the changes configuration. Source row identity is preserved there, along with the original embedding document index.

All unchanged rows retain their original fields. For repaired rows, text is edited and page_ends are remapped (or null if original offsets were invalid). The upstream token_count becomes null: its original value is preserved in original_metadata_json. cleaned_content_tokens in the change manifest is measured using the pinned EmbeddingGemma 2 tokenizer and original prompt convention. Other upstream quality/language metadata describe the original extraction and are retained without recomputation. The original metadata of removed rows is also saved in the change manifest; original text remains in the pinned source.

Existing embeddings are valid for unchanged text. Repaired documents are marked needs_reembedding; this dataset does not silently substitute their old vectors as embeddings of the edited text. No new embeddings are included here.

Processing and publication checkpoint every completed shard batch. An incomplete dataset must not be treated as the finished copy. See _checkpoints/cleanup.json and cleanup-contract.json. Source licensing/provenance are inherited; this derivative does not relicense underlying documents.

Completed results

Retained 8,773,827 documents. Removed 6,567 advertising/spam documents containing 70,737,997 content tokens. Repaired 933 documents, removing 14,941,743 repeated content tokens. Total retained content: 154,769,428,502 EmbeddingGemma tokens.

The final stricter repetition rule was supported in 64/64 reviewed random candidates (57 high confidence, seven medium); this is excerpt-based validation, not a guarantee for every unseen case. Review evidence, uncalibrated embedding ad-ranking scores, and the fitted linear model are in audit/. The complete reversible action plan is cleanup-plan.jsonl. Original text is recoverable from the pinned source.

Final confirmed cleanup

The completed 32-reviewer re-audit added 47 individually confirmed removals (623,836 content tokens). It did not apply new whole-cluster cuts, classifier score thresholds or optional downsampling. Mixed and useful neighboring records remain. Final corpus: 8,773,780 documents, 154,768,804,666 content tokens. The existing 933 repetition repairs remain flagged for re-embedding. Additional IDs, source hashes, original metadata and evidence are retained in changes and audit/reaudit/additional-plan.jsonl.