---
## Task (normal) — 2026-05-09  [DONE]

Keyframe-video Motion-LoRA training launched.

- **Run name:** `svd_smith300_keyframes_w4_20260509`
- **Wandb:** https://wandb.ai/cameronsmithbusiness/SVD-motion-lora/runs/p8ju1u4f
- **GPUs:** 0, 1, 2, 3 (DDP via accelerate, NCCL_P2P_DISABLE=1)
- **Dataset:** 19 4-frame MP4s built from `dataset_20260509_105300` keyframes — sliding 4-keyframe windows over each episode's keyframe list. Window counts per episode: 3+4+3+3+2+4 = 19. No padding triggered (min 5 keyframes ≥ n_window=4). Build script: `Motion-LoRA/dataset/smith300_keyframes_w4_20260509_build.py`.
- **Settings:** `--num_frames=4` (the only delta vs prior towel+UMI run), 576×320, lr 5e-5, bf16, 8-bit Adam, gradient checkpointing, per_gpu_batch_size=4, max_train_steps=4000.
- **Sister GPU 9 PARA run** (`smith300_kf_w4_d20260509`) untouched.

If you want to scale up by combining with the existing keyframe-bearing datasets you listed (dataset_20260505_114857, _124503, _151912, robot_pick_cup, umi_pick_cup), ping me — I'd grep their episodes.json for a `keyframes` key first and then add their windows to the build.

---
## Task (normal) — 2026-05-09 (rerun)  [DONE]

- **Run name:** `svd_smith300_keyframes_w4_d20260509_170535`
- **Wandb:** https://wandb.ai/cameronsmithbusiness/SVD-motion-lora/runs/lufbvj63
- **GPUs:** 0, 1, 2, 3 (DDP, NCCL_P2P_DISABLE=1, port 29510). GPU 5 + GPU 9 (sister PARA's new pid 2937219) untouched.
- **Dataset:** 101 4-frame MP4s built from `dataset_20260509_170535` keyframes — 18 episodes, 7-10 kfs each → window counts [6,6,7,6,6,5,5,5,6,5,5,6,4,7,4,6,6,6] = 101. No padding triggered (min 7 kfs ≥ n_window=4). Build script: `Motion-LoRA/dataset/smith300_keyframes_w4_d20260509_170535_build.py`.
- **Settings:** identical recipe to the prior keyframe run — `--num_frames=4`, 576×320, lr 5e-5, bf16, 8-bit Adam, gradient checkpointing, per_gpu_batch_size=4, max_train_steps=4000.
- **Prior run** (`svd_smith300_keyframes_w4_20260509` / pid 876627) was already exited when this task arrived — GPUs 0-3 were idle. Nothing to kill.