Dataset Viewer
The dataset viewer is not available because its heuristics could not detect any supported data files. You can try uploading some data files, or configuring the data files location manually.
tinygpt-v283 compiled data (v11)
Pre-tokenized SmolLM2-135M (vocab 49152) uint16 token shards.
Mixture
- fineweb-edu-dedup: 79%
- cosmopedia-v2: 10%
- wikipedia: 6%
- smoltalk-instruct: 5%
Total pretrain tokens: 6.40B Total SFT examples: 600001
Filters (v11)
- Exact full-text dedup (blake2b-64)
- MinHash LSH (128 hashes, 16x8 bands, Jaccard 0.85)
- Inter-document line boilerplate (rolling 200K window)
- Wikipedia doc filter + tail cut
- FineWeb-Edu score >= 3.0
- GopherQualityFilter (loosened)
- FineWebQualityFilter (loosened)
- Profanity (drop on 3+ distinct hits)
- Email PII (drop on >5 hits)
- URL blocklist (drop on 3+ hits)
- 100K char doc cap
Removed from v10 for throughput: GopherRepetitionFilter (sources are pre-filtered; drop reasons were not top-10).
Inline strip
Web-boilerplate phrases (see more, read more, see also, click here, subscribe now, etc.) are stripped from each document instead of dropping the doc.
Sources
- HuggingFaceTB/smollm-corpus (fineweb-edu-dedup, cosmopedia-v2)
- wikimedia/wikipedia (20231101.en)
- HuggingFaceTB/smoltalk (data/all/)
- teknium/OpenHermes-2.5 (refs/convert/parquet)
- Downloads last month
- 1,419