Ba2han/merged_pt_2404

Merged pre-training corpus assembled on 2026-04-25 from 7 source datasets. Only the input_ids column is stored, cast to uint16 for storage efficiency.

⚠️ uint16 has a maximum value of 65535. Token IDs above this threshold will overflow. Switch CAST_DTYPE in the build script to uint16 or uint32 if your vocabulary is larger.

Token counts by source

Source dataset	Tokens	Share	Cumulative
tokenized-corpus-0903	5,216,849,173	14.73%	5,216,849,173
warmup_turmix	1,771,460,711	5.00%	6,988,309,884
mixed_corpus-2303	14,211,625,068	40.13%	21,199,934,952
English_mix-tokenized	1,027,372,236	2.90%	22,227,307,188
forumsoh_tokenized	4,232,493,403	11.95%	26,459,800,591
3.5b-tokens	2,879,132,572	8.13%	29,338,933,163
long-corpus	6,072,626,857	17.15%	35,411,560,020
TOTAL	** 35,411,560,020**	100.00%	** 35,411,560,020**

Summary

Metric	Value
Total tokens	35.41 B (35,411,560,020)
Source datasets	7
Parquet shards	473 (~74.87 M tokens/shard)
Shard format	Zstandard-compressed Parquet
Token dtype	uint16
Column	input_ids (list of uint16)
Build date	2026-04-25

Source datasets

• Ba2han/tokenized-corpus-0903
• Ba2han/warmup_turmix
• Ba2han/mixed_corpus-2303
• Ba2han/English_mix-tokenized
• Ba2han/forumsoh_tokenized
• Ba2han/3.5b-tokens
• Ba2han/long-corpus

Reproduction

pip install datasets huggingface_hub tqdm pyarrow
huggingface-cli login
python merge_and_push.py

License

Inherited from source datasets. Check each source repository for its individual license.