source string | target_tokens int64 | tokens int64 | unique_docs int64 | unique_tokens int64 | docs int64 | repeated_token_share float64 | max_copies int64 | copy_pmf dict | seed int64 |
|---|---|---|---|---|---|---|---|---|---|
/model/experiments/0319_verify-ppt/data/marin/dclm_marin | 22,000,000,000 | 22,000,000,202 | 6,413,070 | 7,829,079,546 | 18,025,778 | 0.644133 | 7 | {
"1": 0.46234956941598476,
"2": 0.13512538516988112,
"3": 0.1202448266787148,
"4": 0.02605583651215973,
"5": 0.014412563095710764,
"6": 0.1514967485892226,
"7": 0.09031507053832619
} | 42 |
Marin DCLM with OLMo3-matched repetition (audit corpus)
Audit corpus for "Synthetic Pre-pretraining Survives Scale, but Not as a Grammatical Prior".
Built from verify-ppt/marin-dclm_marin (tokenized with the SmolLM3 tokenizer, DataTrove tokenized-bytes format)
by repeating documents with the copy-count distribution measured on the OLMo3 data used in our runs
(allenai/dolma3_mix-150B-1025; copies per document capped at 7), and shuffling copies across 128 shards.
Used in place of dclm_marin in the Marin mixture; code and math components are unchanged.
See build_stats.json for the build parameters and token counts.
Script: preprocessing/src/build_marin_rep.py in https://github.com/gucci-j/verify-ppt-at-scale
- Downloads last month
- 40