tokens int64 | tokenizer string | light_clean bool |
|---|---|---|
15,180,948,998 | Qwen/Qwen3-0.6B | true |
CommonCrawl HTML, tokenised
Packed token ids for language-model pretraining on raw web markup (not extracted text).
| tokens | 15.18B |
| shards | 77 |
| tokenizer | Qwen/Qwen3-0.6B |
| dtype | uint32, flat, no header |
| document separator | EOS token id |
| preprocessing | light HTML clean (scripts/styles stripped), capped at 32,768 tokens/page |
Shards are not uniform in size: 00000-00047 hold ~92M tokens each (8k pages per source
WARC file), 00048 onward hold ~370M each (29k pages, no per-file cap). Same crawl, same
pipeline, no overlap — shard count is simply not a proxy for token count.
Reading it
import numpy as np
a = np.memmap("CC-MAIN-2025-08-00000.bin", dtype=np.uint32, mode="r")
Concatenate shards, or sample windows across them; each shard is independent and
document-aligned at EOS boundaries. Every id is < 151937; anything else means the file was
read at the wrong dtype.
Note the model config that consumes this uses vocab_size=151937, while the tokenizer reports
151670. Use the larger value or embedding lookups will be out of range.
Provenance
Pages harvested from CommonCrawl (CC-MAIN-2025-08). Markup is retained deliberately — the intended use is modelling HTML structure, so stripping tags would defeat the purpose.
- Downloads last month
- 106