| # `data/python_splits/` — Python task splits |
| |
| `jsonl` splits and derived files for **verified** code completion / repair benchmarks. |
| |
| ## Core `jsonl` files |
| |
| | File | Description | |
| |------|-------------| |
| | `train_verified.jsonl` | Training set | |
| | `val.jsonl` | Validation set | |
| | `test_verified.jsonl` | Test set | |
| | `train_valtest.jsonl` | Train+val+test union (if a script needs the full pool) | |
| | `codegen_1shot_test.jsonl` | One-shot generation test subset | |
| | `test_input_verified_baseline.jsonl` | Baseline input version | |
| | `test_input_verified_baseline_with_weights.jsonl` | With per-sample weights | |
| | `test_input_verified_baseline_with_cfg_weights.jsonl` | With CFG-related weights | |
| | `train_verified_with_weights.jsonl`, `train_verified_with_cfg_weights.jsonl` | Weighted training | |
| | `test_verified.target_with_token_buckets.jsonl` | Test rows with **target length bucket** labels | |
|
|
| ## Bucket metadata |
|
|
| | Path | Description | |
| |------|-------------| |
| | `test_verified.target_bucket_subsets/` | Per-bucket files, e.g. `1_128.jsonl`, `128_256.jsonl`, `256_512.jsonl`, `512_1024.jsonl` | |
| | `test_verified.target_bucket_to_keys.json` | Bucket name → list of sample ids | |
| | `test_verified.target_token_bucket_summary.json` | Per-bucket counts and stats | |
|
|
| `rl.py` function `_resolve_train_subset_specs` reads selected files under `python_splits/test_verified.target_bucket_subsets` by default and infers per-bucket `max_new_tokens`, `steps`, and `shrink` (overridable in YAML). |
|
|
| ## Other |
|
|
| | File | Description | |
| |------|-------------| |
| | `command.txt` | Commands used to generate or refresh this folder | |
|
|
