dit_block_tower_norm_fix

Diffusion policy checkpoint for the build_block_tower task (with DAgger rounds 1.0.0–1.4.0), trained with per-timestep (H,D) RAMEN action normalization and corrected action chunk semantics (slot 0 = current action).

Checkpoint

Step	train_loss	Status	Hash
29000	0.0047	partial (29k/50k)	8843965c8dcf0fc68b71784fe0875b7e43eb25aa4e05b59bf796b2280b50ea96

Training was interrupted by walltime (1 day limit) at step ~29588/50000. Loss was still decreasing healthily — this checkpoint will be resumed.

Config

• dataset: villekuosmanen/build_block_tower + DAgger rounds 1.0.0–1.4.0
• batch_size: 80 per GPU (320 global, 4x GPUs)
• optimizer_lr: 3e-4
• lr_scheduler: cosine (warmup 500 steps, min_lr_scale 0.1)
• horizon: 32, n_action_steps: 32
• noise_scheduler: DDIM, 100 train timesteps, 20 inference steps
• observation_encoder: CLIP ViT-B/16 (vision + text)
• action normalization: RAMEN with per-timestep stats (H=32, D=17)

Files

checkpoints/29000/params/model.safetensors   # model weights
checkpoints/29000/params/config.json         # model config
assets/ramen_stats.json                      # action normalization stats
TRAINING_LOG.md                              # sanitized training log

Verify integrity

cd checkpoints/29000/params
find config.json model.safetensors -type f | sort | xargs sha256sum | sha256sum
# expected: 8843965c8dcf0fc68b71784fe0875b7e43eb25aa4e05b59bf796b2280b50ea96

Note: ramen_stats.json is in assets/, not in the params directory. The hash above covers only the params files. To reproduce the full hash including ramen_stats, download all three files into one directory and run the same command over all three.

W&B

• Training dashboard

Source

• repo: pravsels/multitask_dit_policy (branch stage1-multimodal-abstraction)