Skip to content

Published SocialNav checkpoints appear to be unfinished mid-training snapshots #10

Description

@danielliujm

Hi, Thank you for this great work

While trying to run flow-matching waypoint inference against the released checkpoints (Qwen2.5-VL-SAFE-GRPO and Qwen2-VL-SAFE-GRPO via ModelScope, following the train_branch="fm" path from utils/citywalker.py), the predicted waypoints don't appear to be sensible on real test image sequences, not resembling plausible navigation trajectories.

The checkpoint files themselves might suggest why, as both published checkpoints look like unfinished, mid-training GRPO snapshots.

  • trainer_state.json shows training appears to be stopped early in both cases:

    • Qwen2.5-VL-SAFE-GRPO: global_step=8000 / max_steps=125000 (6.4%), epoch=0.32/5
    • Qwen2-VL-SAFE-GRPO: global_step=17600 / max_steps=301420 (5.8%), epoch=0.58/10
  • Both ship full distributed-training resumption state (optimizer.pt, scheduler.pt, training_args.bin, per-rank rng_state_N.pth
    files)

  • best_global_step/best_metric/best_model_checkpoint are all null in both trainer_state.json files.

  • The HF contain only the auto-generated .gitattributes file.

Question: are the ModelScope uploads intended to be the final checkpoints, or were the fully-trained weights not yet pushed? If
these are final, are there any known issues with the flow-matching action head at inference time that would explain nonsensical
waypoint output?

Thank you so much for your help.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions