# `data/preprocessed/sft_dream_py_ast/` — AST SFT caches (many runs)

Sibling to `sft_dream_py/`. Each **subfolder** is one **AST-related preprocessing** pipeline (mask policy, depth, random depth, SVEN subsets, etc.), usually containing `train.pt`, `val.pt`, `meta.json` (verify on disk).

## Subfolder name patterns (examples)

| Pattern | Meaning (high level) |
|---------|----------------------|
| `ast_depth_masking*`, `ast_depth_masking_both`, `ast_depth_masking_min` | Mask variants by **AST depth** |
| `ast_depth_masking_invert*` | Inverted or ablation depth masks |
| `random_depth_masking*` | **Random depth** masks |
| `ast_node_masking_*` | **Node-level** masks |
| `ast_random_m_*` | Random-AST `m` experiment series |
| `*_sven_1_1024`, `*_sven_bigvul*` | **SVEN / Big-Vul** scale or subset |
| `ast_depth_label_learning` | Preprocessing with **depth label learning** |

We **do not** add a README in every leaf; for reproduction use that folder’s `meta.json` and the YAML `dataset.preprocessed_dir` from the run.