data/preprocessed/ — Preprocessed tensor caches
Tokenize jsonl offline into .pt bundles to reduce CPU load during SFT or related stages.
Subdirectories
sft_dream_py/
| File | Description |
|---|---|
train.pt |
Training split tensor dict (input_ids, labels, attention_mask, etc.; confirm with torch.load) |
val.pt |
Validation split |
meta.json |
Source paths, tokenizer name, max length, etc. |
sft_dream_py_ast/ (if present)
Same general layout with AST / extra fields; often contains many experiment subfolders—see sft_dream_py_ast/README.md. Tensor keys follow train/sft_dream_dataset_ast*.py.
Generation is usually a custom script or a --preprocess-style path in train/; dataset.preprocessed_dir in YAML should match the folder name.