SIP-nesting

SIP evaluation checkpoint for nesting Russian dolls, initialized from Long-WAM-YAM.

Task

Nest the Russian dolls from smallest to largest by placing each smaller doll inside the next larger one, repeating until all dolls are nested securely inside the largest doll.

Files

• model.pt: original evaluation checkpoint, uploaded without modifying its weights or embedded metadata.
• config.yaml: model, observation, action, normalization and inference settings; local paths are portable placeholders.
• dataset_stats.json: the normalization statistics used by this run.
• requirements-sip.txt: matching SIP environment package versions; install the matching Long-WAM source checkout at ./repo.

Inference

Use the SIP Long-WAM inference implementation with this repository's model.pt, config.yaml and dataset_stats.json. The matching source release is retained separately with the SIP project. Load the config with longwam.inference.load_run_config, then pass the explicit model.pt path to longwam.inference.load_model.

The configuration uses BF16, three camera views (top_head, hand_left, hand_right), a 384 x 320 combined layout, P48 history, a 32-step action chunk, and 14D absolute joint/gripper targets. Preserve the original gripper units and task-specific z-score normalization without clipping. Adapt relative dataset, statistics and optional text-cache paths to your inference machine. Wan VAE and UMT5 tokenizer/text-encoder assets are required separately.

This is an evaluation checkpoint. It does not contain complete optimizer, scheduler or random state for training resumption. No robot success rate or held-out performance is claimed.

Base model: https://huggingface.co/Efficient-Large-Model/Long-WAM-YAM