metadata
license: mit
Dataset
Mixed pretraining dataset built from:
| Source | Weight | Notes |
|---|---|---|
| epfml/FineWeb-HQ | 60% | Quality-filtered web text (score > 0.8) |
| HuggingFaceTB/cosmopedia | 20% | Sequential: stanford → wikihow → web_samples_v2 |
| HuggingFaceTB/finemath (finemath-4plus) | 10% | Mathematical reasoning |
| bigcode/python-stack-v1-functions-filtered | 10% | Python code functions |
Total rows: 2,000,000
Shards: 20 parquet files
Split: all rows are in train
Around - 1.59B Tokens