Robotics
GWM / real_data /README.md
Shady0057's picture
Add real-data GWM checkpoint at step 34000
ec94a57 verified
|
Raw History Blame Contribute Delete
2.62 kB

GWM real-data checkpoint (step 34000)

This is the GWM checkpoint referenced by the real-robot experiment code as 0810_gwm step 34000. It was trained on a mixture of MolmoAct2-DROID real robot data and MolmoBot simulation data using the GWM training pipeline.

Source code: https://github.com/QuanyiLi/gwm-wiser

  • Training step: 34,000 (latest saved checkpoint in this run).
  • Weight format: canonical PyTorch checkpoint.pt, with BF16 model weights, model configuration, training metadata, and dependency versions.
  • File size: 1,493,324,789 bytes (1.49 GB).
  • Architecture: 5 transformer layers, width 4096, FFN width 8192, 32 attention heads, 8 KV heads; input/output features use 1620 x 4096.
  • Feature encoder: Qwen/Qwen3-VL-Embedding-8B (download separately).
  • Saved with PyTorch 2.7.1+cu126 and Transformers 4.57.6.

The original training log reports held-out open-loop MSE 0.01902 and cosine similarity 0.8457 at step 34,000, over 8 evaluation batches. These are logged training-time metrics, not a new evaluation of this upload. The run planned 37,500 steps but was cancelled after step 34,860; the most recent saved checkpoint is step 34,000.

Download

hf download Shady0057/GWM real_data/checkpoint.pt --local-dir gwm_ckpt

The checkpoint will be at gwm_ckpt/real_data/checkpoint.pt.

Load

Run from a checkout of the source repository with its dependencies installed:

from huggingface_hub import hf_hub_download
from real_data_train.gwm_model import load_canonical_like_planner

path = hf_hub_download("Shady0057/GWM", "real_data/checkpoint.pt")
model, checkpoint = load_canonical_like_planner(path)
model.eval()
print(checkpoint["step"])  # 34000

The checkpoint contains a serialized TransformerConfig from gwm_wiser.models.transformer, so the source repository must be importable. The loader reconstructs the model and strictly loads its state dictionary. The companion Qwen encoder is required for the complete inference pipeline.

Provenance and verification

training_config.json preserves the original resolved training configuration; its local filesystem paths describe the training machine and should be adapted when reproducing the run. checkpoint_metadata.json records the architecture, logged evaluation metrics, and validation results. SHA256SUMS records the checksum of the unmodified checkpoint.

Before upload, all state dictionary keys and shapes matched the canonical model, and all floating-point tensor values were finite. This directory contains inference weights; optimizer/resume state is not included.