Think-Dataset (Filtered from Institutional Books)

Filters

Field	Rule
language_gen	== "eng"
language_distribution_gen	English proportion >= 0.9
date1_src → date2_src fallback	parsed year < 1930 (undated REJECTED; invalid date types rejected: True)
ocr_score_src	>= 90.0
ocr_score_gen	>= 90.0
OCR agreement	|src - gen| <= 10.0
text_analysis_gen.text_by_page_gen.tokenizability_score	>= 95.0
tiny fragments	tokens >= 500, chars >= 2000, pages >= 3, sentences >= 20 unless token exception applies

Dedup

Enabled — via likely_duplicates_barcodes_gen barcode claiming (first-seen representative kept).

Diversity strategy

• Single shuffle buffer: Shards 00000-00449 were produced with a 20 GB shuffle buffer. The final 23 shards, 00450-00472, used an 8 GB buffer to reduce Colab memory pressure. Books were randomly evicted after the active buffer exceeded its limit, reducing source-order clustering while preserving the corpus's natural topic distribution.
• Validation split: deterministic by barcode hash — a book is val iff crc32(barcode) % 370 == 0 (~`1000` books). Resume-stable; written as the last lexicographic shard.

Stats

• Source rows seen: 983,004
• Rows passed filter + dedup: 202,470 (20.60% of seen)
• Total characters written: 118,745,375,871
• Total shards: 473 (train: 472, val: 1)
• Last shard (val): shard_00472.parquet

Rejections by reason (not accurate)

• date_type_invalid: 235,713
• duplicate: 35,229
• language: 1,172,869
• language_distribution_missing: 3,059
• language_low_english_proportion: 207,550
• ocr_low: 209,112
• ocr_missing: 3
• tokenizability_low: 45,564
• tokenizability_missing: 12
• too_few_sentences: 3
• year_too_recent: 22,962
• year_unparseable: 612

Estimated language filtering

Exact rejection totals were not preserved because interrupted resumes counted some rejected source rows more than once. A clean 50,000-row dry run using the final filters rejected 18,851 non-English books, 8,772 books below the 90% English threshold, and 9 with missing language metadata. Extrapolated across 983,004 source rows, approximately 543,200 books (55.3%) were removed by language filtering. These figures are estimates, not exact counts.

Pass-by-year-source (date1_src vs date2_src fallback recovery)

• date1_src: 202,458
• date2_src: 12

Overall topic distribution

• LANGUAGE AND LITERATURE: 45,863 (28.62%)
• PHILOSOPHY. PSYCHOLOGY. RELIGION: 29,548 (18.44%)
• LAW: 13,798 (8.61%)
• SCIENCE: 13,308 (8.30%)
• HISTORY OF THE AMERICAS: 9,543 (5.95%)
• SOCIAL SCIENCES: 8,603 (5.37%)
• AUXILIARY SCIENCES OF HISTORY: 5,404 (3.37%)
• AGRICULTURE: 5,290 (3.30%)
• POLITICAL SCIENCE: 4,966 (3.10%)
• EDUCATION: 4,381 (2.73%)
• TECHNOLOGY: 3,733 (2.33%)
• GEOGRAPHY. ANTHROPOLOGY. RECREATION: 3,433 (2.14%)
• FINE ARTS: 3,192 (1.99%)
• MEDICINE: 3,034 (1.89%)
• MUSIC AND BOOKS ON MUSIC: 1,573 (0.98%)
• WORLD HISTORY AND HISTORY OF EUROPE, ASIA, AFRICA, AUSTRALIA, NEW ZEALAND, ETC.: 1,481 (0.92%)
• NAVAL SCIENCE: 1,000 (0.62%)
• GENERAL WORKS: 914 (0.57%)
• MILITARY SCIENCE: 829 (0.52%)
• BIBLIOGRAPHY. LIBRARY SCIENCE. INFORMATION RESOURCES (GENERAL): 321 (0.20%)
• UNKNOWN: 49 (0.03%)

Schema

Single string column named text. ZSTD-3, row-group size 64. Drop-in compatible with nanochat/dataset.py. Per-shard topic distribution lives in manifest.json (shards[].topic_distribution). Per-document audit metadata lives separately in audit_metadata.jsonl; training shards remain text-only.

Use with nanochat

Rename to base_data_climbmix/ (matches DATA_DIR in nanochat/dataset.py:27) or edit that constant to point at this directory. Then re-train the tokenizer:

python -m scripts.tok_train && python -m scripts.tok_eval
python -m scripts.base_train --depth=12 --window-pattern=L