yang29's picture
Document checkpoint formats and verified retirement
7ad9e47 verified
|
Raw History Blame Contribute Delete
1.4 kB
---
tags:
- text-to-video
- self-forcing
- causal-forcing
- astrolabe
- advdmd
---
# AFD baseline checkpoint archive
Evaluated short-video Astrolabe and AdvDMD checkpoints are stored under
`short-self-forcing/METHOD/step-NNN/` and `short-causal-forcing/METHOD/step-NNN/`.
Evaluated long-video compact inference exports are under
`long-self-forcing/METHOD/step-NNN/`. These exports preserve the original saved
format; they are distinct from full restart checkpoints retained by active
long-video training. Short Astrolabe bundles include LoRA configuration,
LoRA weights and the original optimizer/scaler files. Short AdvDMD files
contain the original generator LoRA and discriminator heads.
Every bundle includes `archive_provenance.json`: evaluation provenance,
original local filenames, per-file byte sizes and SHA256 checksums. Local
retirement is authorized only after remote checksum verification at an
immutable Hub revision. Original initialization weights remain in their
upstream repositories and are not included in this archive.
Results and generation/configuration provenance:
- https://huggingface.co/datasets/yang29/afd-baselines-vbench-step80-20261109
- https://huggingface.co/datasets/yang29/afd-baselines-vbench-step160-20261110
- https://huggingface.co/datasets/yang29/afd-long-video-vbench-100prompts
- https://huggingface.co/datasets/yang29/afd-causal-forcing-baselines-vbench