| --- |
| license: apache-2.0 |
| tags: |
| - video-generation |
| - reinforcement-learning |
| - grpo |
| - lora |
| --- |
| |
| # old_checkpoints |
| |
| Archived training logs and LoRA checkpoints from RL fine-tuning (GRPO-family) of |
| video diffusion models. |
| |
| ## Contents |
| |
| ### `skyreels_all.zip` — SkyReels-I2V runs (54.37 GB, split into 6 parts) |
|
|
| 13 experiment trees + `_promptdemo_skyreels`: 24,864 files, 633 symlinks, 2,734 directories. |
| Each run holds `checkpoints/` (LoRA current/old/EMA + optimizer state), `permanent/epoch_N/` |
| snapshots, `evaluation_vbench_i2v/` videos and scores, and `wandb/` logs. |
|
|
| Runs: `skyreels_default` (4 variants: epg / pepg / grpo / epg-guidance5.0), `skyreels_bump`, |
| `skyreels_bump_epg`, `skyreels_bump_ram`, `skyreels_bump_diffusionnft`, `skyreels_bump_ta05_epg`, |
| `skyreels_diffusionnft`, `skyreels_ram`, `skyreels_noise_split_dd08`, `skyreels_flow_grpo`, |
| `skyreels_i2v_base_eval`, `skyreels_wan_extended_a14b`, `_promptdemo_skyreels`. |
|
|
| ## Reassembly |
|
|
| The zip exceeds the Hugging Face 50 GB per-file limit, so it is stored as 10 GB parts. |
|
|
| ```bash |
| huggingface-cli download toony42/old_checkpoints --local-dir old_checkpoints |
| cd old_checkpoints |
| cat skyreels_all.zip.part* > skyreels_all.zip |
| sha256sum -c skyreels_all.zip.sha256 # verify |
| unzip skyreels_all.zip |
| ``` |
|
|
| The archive is **stored, not deflated** (`zip -0`) — the payload is bf16 tensors and mp4 video, |
| which compress by only ~9%, so compression was not worth the CPU cost. |
|
|
| ## Notes |
|
|
| - Symlinks are stored as links (`zip -y`). 590 of them point at `~/.cache/wandb/*` core-debug |
| logs outside the archive and will not resolve after extraction; they are wandb debug logs only. |
| - wandb metadata embeds original absolute paths, hostname, and username from the training cluster. |
|
|
