| # `data/sven_rl_buckets/` — SVEN RL length buckets | |
| Produced by `scripts/convert_sven_to_dream_jsonl.py` (and similar), splitting SVEN examples by **target generation length** so RL rollout can set per-bucket `max_new_tokens` / `steps`. | |
| ## Layout | |
| ``` | |
| train/ | |
| 1_128.jsonl | |
| 128_256.jsonl | |
| 256_512.jsonl | |
| 512_inf.jsonl # or similar; 512+ tokens | |
| val/ | |
| (same pattern) | |
| ``` | |
| `rl_sven.py` and config keys `dataset.sven_train_subset_dir`, `dataset.sven_train_subset_files` point here; semantics mirror `python_splits` bucketing. | |