text stringlengths 2.91k 11.4M |
|---|
"CHAPTER V.\n\n[1562 to 1566.]\n\nProgress of the Troubles in the Low Countries, p. 1.-Cardinal Gran(...TRUNCATED) |
"BAR ASSOCIATION OF SAN FRANCISCO\n\nUS 908.3CA\n\nBAR\n\nBAR ASSOCIATION OF SAN FRANCISCO \"The Unl(...TRUNCATED) |
"INTRODUCTION\n\ncian should be particularly indebted to the author, as her book will be a valuable (...TRUNCATED) |
"CHAPTER I.\n\nAncient Brewster families\n\nWilliam Brewster's probable relationship\n\nTime of his (...TRUNCATED) |
"PART II. SCRIPTURE LESSONS.\n\nI-Pericopee........\n\nPAGE II-Penitential Psalms.\n\nIII.History of(...TRUNCATED) |
"Tharping THE FOOD OF LONDON.\n\nLONDON:\n\nPrinted by SPOTTISWOODE & Co.\n\nNew-street-Square\n\nCO(...TRUNCATED) |
"COLLECTED PAPERS ON ACOUSTICS\n\n14 898 78 QC22 доба\n\nBY WALLACE CLEMENT SABINE LATE HOLLIS P(...TRUNCATED) |
"CHAPTER I THE PRIEST II THE BEGGAR\n\nPART I\n\nI THE LANDSCAPE II CHRISTMAS EVE III CHRISTMAS DAY\(...TRUNCATED) |
"PREFACE\n\nNOR the past thirteen years there have been submitted to the General Counsel by members (...TRUNCATED) |
"PREFACE\n\nThe arrangement and presentation of the material in this volume are the same as in Volum(...TRUNCATED) |
End of preview. Expand in Data Studio
think-midtrain
Midtraining corpus assembled from pre-1930s public-domain STEM-leaning text, built in five
sequential stages. Each stage reads the previous stage's output from this repo and writes its own
prefix; every stage keeps whole documents as single rows and preserves per-document metadata
(title, OCR scores, token counts, source identifiers, ...) alongside text.
Pipeline stages
| Stage | Prefix | Notebook | What it does |
|---|---|---|---|
| 1 | (external: jbduran/think-dataset) |
midtraining_extraction.ipynb |
LC-classification + OCR/tokenizability filtering of jbduran/think-dataset, targeting underrepresented Science/Technology/Medicine subjects. |
| 2 | stage2/ |
midtraining_extraction.ipynb |
Internet Archive harvest (math + STEM collections, pre-1930 date range), heuristic OCR-confidence + language + dedup filtering. |
| 3 | stage3/ |
midtraining_ocroscope_filtering.ipynb |
Second-pass OCR quality filter using Pleias/OCRoscope (7-gram language-consistency + non-character ratio), independent of Stage 2's dictionary-hit heuristic. |
| 4 | stage4/ |
midtraining_stage4_cleaning.ipynb |
Boilerplate/OCR cleanup (Google Books, HathiTrust, library stamps, Project Gutenberg headers, hyphenation, front matter) + a GPT-2 token log-prior filter (p2.5-p97.5) pruning statistical outliers. Port of think-dataset-clean. |
| 5 | stage5/ |
midtraining_stage4_cleaning.ipynb |
Footer/boilerplate line stripping, then a tiered 1930-cutoff anachronism keyword filter (regex/keyword matcher; a document is dropped only on strong evidence of post-1930 content). Port of think-dataset-clean-1930s. |
Current totals (as of the last manifest update on each prefix)
- Stage 2 (Internet Archive harvest): 15,075 documents, 1,622,288,069 tokens (o200k_base)
- Stage 3 (OCRoscope filter): 15,075 in -> 13,552 kept
- Stage 4 (structural clean + log-prior filter): 13,552 in -> 12,791 kept
- Stage 5 (footer strip + anachronism filter): 12,791 in -> 11,409 kept
Resumability
Every stage is resumable at shard granularity: a rerun skips any output shard already present on
HF and continues with whatever remains. Stage 4's GPT-2 log-prior table and Stage 5's banned-term
list are each built once, cached under stage4/_prior/ and stage5/_banned/, and reused on
every subsequent run instead of being recomputed.
Artifacts per stage
stage2/:data/,audit_metadata.jsonl,checkpoint.jsonl,manifest.jsonstage3/:data/,scored/(OCRoscope score checkpoints),manifest.jsonstage4/:data/,stats/,_prior/(cached log-prior table + thresholds),manifest.jsonstage5/:stripped/,strip_stats/,strip_samples/,data/,stats/,hits/,_banned/(cached term list + tiers),cleaning_report_1930s.json,manifest.json
Updated: 2026-07-20T19:08:37Z
- Downloads last month
- 750