| license: odc-by | |
| task_categories: | |
| - text-generation | |
| language: | |
| - en | |
| # dolma3_6T_sample_50000_docs | |
| Materialized working sample from the Dolma3 6T deduplicated corpus. | |
| ## Parameters | |
| | Parameter | Value | | |
| |-----------|-------| | |
| | Sampling mode | 50,000 docs/bin | | |
| | Seed | 42 | | |
| ## Counts | |
| | Metric | Value | | |
| |--------|-------| | |
| | Total documents | 26,249,124 | | |
| | Total tokens | 62,819,501,017 | | |
| | Bins covered | 576/576 | | |
| | Bins underfilled | 79 | | |
| ## Format | |
| JSONL files compressed with zstandard under `data/`. | |
| Each record contains the original Dolma document fields (id, text, metadata). | |
| Sampling metadata files: | |
| - `working_sample_manifest.parquet` - sampled doc_ids with bin assignments | |
| - `sample_contract.json` - aggregate sampling contract | |
| - `bin_summary.csv` - per-bin fill rates (576 bins = 24 topics x 24 formats) | |