| # Tier C: Exact Retraining Data | |
| This directory contains the exact local training inputs used by the released | |
| 1M Stage-I and 300K SFT experiments. | |
| ```text | |
| stage1_1m/recap_train_1m/ 200 WebDataset tar files | |
| sft_300k/ 64 packed parquet shards | |
| ``` | |
| The Stage-I tar files are sparse on the source filesystem: their allocated | |
| disk usage is smaller than their logical byte size. The manifests record the | |
| logical size and SHA-256 of the complete file stream, which is the relevant | |
| size after download. | |
| Excluded as redundant: | |
| - `extra_800k_chunks`: construction intermediates already represented by the | |
| final 200 Stage-I tar files. | |
| - raw SFT sources: the released 64 packed parquet shards are the exact inputs | |
| consumed by the training code. | |
| See the GitHub training configurations and scripts for the original relative | |
| paths and optimizer settings. | |