| # Dataset layout | |
| All data files are stored physically under `dataset/`. There are no symlinks. | |
| | Path | Meaning | Consumed by | | |
| |---|---|---| | |
| | `dataset/source_csv/` | Original clustered CSV traces for 20 devices | pair/outpaint precompute and raw evaluation | | |
| | `dataset/windows_v5/` | Time-ordered 256-token windows before V2 tokenization | V2 conversion and condition reconstruction | | |
| | `dataset/tokenized_v2/` | VAE-ready DELTA-256 + RETURN-32 representation | VAE training and decoding | | |
| | `dataset/precomputed/singlegen/` | VAE latents plus 7-D conditions for individual windows | single-window DiT | | |
| | `dataset/precomputed/outgen/` | Prefix/suffix latent pairs plus 14-D conditions | outpaint DiT | | |
| The flow is: | |
| ```text | |
| source_csv ────────────────┐ | |
| ├─> precomputed/outgen ─> checkpoints/outgen | |
| windows_v5 ─> tokenized_v2┤ | |
| └─> precomputed/singlegen ─> checkpoints/singlegen | |
| └────────> checkpoints/vae | |
| ``` | |
| `precomputed` is still data, not a model checkpoint. It contains frozen VAE | |
| latent tensors used to avoid encoding all windows again during DiT training. | |
| The actual learned weights live only under `checkpoints/`. | |