# `data/` — Datasets and indices

Data for **Python code generation / repair**: HumanEval-style `jsonl`, length-bucketed subsets, SVEN-related json/jsonl, and optional **torch-preprocessed** caches. Often large and **gitignored**; fetch or generate separately.

## Top-level files (examples)

| Path | Description |
|------|-------------|
| `sven_rl_train.jsonl`, `sven_rl_val.jsonl` | Merged SVEN RL lists (combine with `sven_rl_buckets/` when using buckets) |
| `sven_bigvul_sft_all_1_1024_{train,val}.json` | Big-Vul subset SFT splits |

## Subdirectories

| Path | Description |
|------|-------------|
| [`python_splits/`](python_splits/README.md) | Train/val/test `jsonl`, weighted variants, **target-token buckets** |
| [`preprocessed/`](preprocessed/README.md) | `train.pt` / `val.pt` / `meta.json` tokenized caches |
| [`sven_rl_buckets/`](sven_rl_buckets/README.md) | **Length-bucketed** `jsonl` for RL (`train/`, `val/`) |
| [`codenet/`](codenet/README.md) | CodeNet **public test cases** and related files |

## Preparation tips

- **Buckets**: Keep `scripts/classify_target_tokens_dream.py` consistent with `python_splits/test_verified.target_bucket_subsets/`.
- **SVEN RL**: Prefer `scripts/convert_sven_to_dream_jsonl.py` to populate `sven_rl_buckets/`.