README.md

metadata

configs:
  - config_name: default
    data_files:
      - split: temiz_wiki
        path: data/temiz_wiki-*.parquet
      - split: nemotron_economics
        path: data/nemotron_economics-*.parquet
      - split: nemotron_formal_logic
        path: data/nemotron_formal_logic-*.parquet
      - split: nemotron_unconditional_algorithmic
        path: data/nemotron_unconditional_algorithmic-*.parquet
      - split: assistant_answers
        path: data/assistant_answers-*.parquet
      - split: augmented_0303
        path: data/augmented_0303-*.parquet
      - split: augmented_all
        path: data/augmented_all-*.parquet
      - split: curated_mixed
        path: data/curated_mixed-*.parquet
      - split: long_3005
        path: data/long_3005-*.parquet
      - split: cleaned_ocr
        path: data/cleaned_ocr-*.parquet
      - split: long_1811
        path: data/long_1811-*.parquet

Ba2han/1907-long

Long-form pretraining data tokenized with Ba2han/outputs.

• Character filter: 1,500 to 15,000 characters, inclusive
• Every sequence is wrapped once with the tokenizer's BOS and EOS IDs
• Stored columns: input_ids (list<int32>) and token_count (int32)
• Each split is shuffled deterministically before export
• Parquet shard size: up to 200,000 rows

Token statistics

Split	Rows	Average	Median	Min	Max	Total tokens	Shards
temiz_wiki	111,482	853.19	622	245	8,054	95,115,014	1
nemotron_economics	58,253	704.78	668	286	4,234	41,055,788	1
nemotron_formal_logic	80,523	952.11	905	342	5,354	76,666,921	1
nemotron_unconditional_algorithmic	133,493	1,980.53	1,696	412	10,904	264,386,959	1
assistant_answers	216,843	906.70	857	253	5,249	196,612,346	2
augmented_0303	2,105,912	662.75	586	191	7,077	1,395,697,703	11
augmented_all	122,462	586.36	520	229	7,071	71,807,070	1
curated_mixed	68,643	759.71	702	246	6,413	52,148,500	1
long_3005	530,872	991.19	779	191	11,794	526,192,772	3
cleaned_ocr	8,363	1,285.56	1,014	296	10,803	10,751,129	1
long_1811	33,482	1,403.36	1,558	262	5,260	46,987,272	1
All	3,470,328	800.33	661	191	11,794	2,777,421,474	24

Source licenses and attribution requirements remain applicable to their respective splits.