The source dataset mixes reconstructed chunk sequences and singleton chunk rows. For this unchunked release, rows were grouped by metadata.id; when duplicated singleton rows were present for the same document, the reconstruction kept the series with the largest chunk_total. Text was joined with inferred text overlap.

Reconstruction Summary

Source: costadev00/gutenberg-project-tokenweaver-cpt-2048
Source revision: 8ebd9290e1faee021ed9d7cd7e994769adfb5330
Source config/split: default/train
Strategy: text-overlap
Group id column: metadata.id
Max overlap: 4096 characters
Prefer max chunk_total: true
Input rows: 40152
Candidate groups: 647
Reconstructed groups: 638
Quarantined groups: 9

Files

data/train.parquet: reconstructed full-document rows.
reports/reconstruction_report.json: full reconstruction report.
reports/quarantined_groups.jsonl: groups skipped by the safety checks.

The quarantined groups were not included in data/train.parquet.

Downloads last month: -

Total file size:

65.8 MB