| --- |
| pretty_name: Distillation assets |
| tags: |
| - video |
| - distillation |
| --- |
| # Distillation assets |
|
|
| Processed training-data tensors, model weights and an EMA checkpoint. |
| Only a selected subset of the dataset is included. The full metadata ordering |
| is retained; files absent from this subset must be obtained separately for |
| other training schedules. |
|
|
| ## Download and restore |
|
|
| Downloads require approved access. Request access on the dataset page and sign |
| in with an approved account before downloading. |
|
|
| ```bash |
| python3 -m pip install -U huggingface_hub zstandard |
| hf auth login |
| hf download mingyang-wu/Distillation --repo-type dataset \ |
| --include 'assets-v1/*' --local-dir ./asset-download |
| python3 ./asset-download/assets-v1/restore.py --dest ./project |
| ``` |
|
|
| ```text |
| project/ |
| ├── data/ |
| ├── models/base-transformer/ |
| ├── checkpoints/step-2100/ema/ |
| └── metadata/training-manifest.json |
| ``` |
|
|
| Archive streams are split into 2 GiB parts. Keep all parts together. |
| The restore script verifies archive and individual-file SHA256 checksums. |
| Existing files at matching destination paths are replaced. |
| Metadata file paths are relative to the restored project directory. |
|
|
| Tensor payloads are preserved. Numerical model settings are included without |
| automatic class-dispatch or source-location metadata. Model implementations, |
| training code, public model components and runtime dependencies are supplied |
| separately; this asset repository does not install or launch training. |
|
|
| ## Filtered data: assets-v2 |
|
|
| The separate assets-v2 bundle contains 6,400 quality-filtered latent/text pairs for a fixed 200-update schedule. Its full manifest retains 36,058 records. |
|
|
| ```bash |
| hf download mingyang-wu/Distillation --repo-type dataset \ |
| --include 'assets-v2/*' --local-dir ./asset-download |
| python3 ./asset-download/assets-v2/restore.py --dest ./project \ |
| --on-conflict keep --delete-archives |
| ``` |
|
|
| V2 uses the same `--dest` as v1. By default its data is restored into the shared |
| `data/` tree, so v1 and v2 manifests use the same relative paths; the v2 manifest |
| is written to `metadata/training-manifest-v2.json`. Use `--dry-run` to preview, |
| `--on-conflict keep` when v1 is already restored, and `--delete-archives` to |
| remove the parts after a verified restore. Model weights remain in assets-v1. |
| See [assets-v2/README.md](assets-v2/README.md) for all options, the layout and |
| the sampling contract. |
|
|
