Missing released BDD-X PGM checkpoint (pgm/ckpts/pgm/bddx_weights.npy)

[jzstark]

Hello SafeAuto authors,

I am reproducing the released BDD-X evaluation:
bash scripts/eval_bddx.sh 0.01 0.35 true 2

The evaluator requires pgm/ckpts/pgm/bddx_weights.npy. However, this checkpoint is not present in the GitHub repository or the released Hugging Face SafeAuto-BDDX model files.

The released pgm/predicates/bddx/train_vectors.pkl is 36-dimensional, while the BDDX MLN rules index the 52-dimensional predicate space, so python -m pgm.pgm --dataset bddx fails with an IndexError. Could you please release the BDD-X PGM checkpoint, or provide:

1. a download URL for bddx_weights.npy,
2. its SHA-256 checksum,
3. the exact commit/dataset revision and training command used to produce it,
4. clarification on how the 52-dimensional training predicates were generated?

Thank you!

[javyduck]

Hi, thanks for the interest in our work and for the very careful digging — sorry for the missing checkpoint! Everything should be resolved on the latest main:

1. Checkpoint released: the original BDD-X PGM checkpoint used in the paper is now included in the repo at pgm/ckpts/pgm/bddx_weights.npy (sha256: f142fae50c0774c26985f6860eb8c0a77f558b25130418b6e06d449eaed9ed86), so bash scripts/eval_bddx.sh 0.01 0.35 true 2 works out of the box.
2. 36-dim vs 52-dim: the released train_vectors.pkl contains only the observed predicates (16 actions + 12 environment + 8 control signals = 36 dims). The original checkpoint was trained with the 16 MLLM action predicates additionally included — they are generated from the MLLM's predicted actions on the training set, giving the full 52-dim space.
3. Retraining with the released vectors: python -m pgm.pgm --dataset bddx now works — pgm/pgm.py automatically zero-pads the missing MLLM block (those rules are trivially satisfied during training and only take effect at inference), so the IndexError is gone. The training loop is also vectorized now and finishes in seconds. We additionally fixed a PGM(weights_path=...) typo in eval_custom_predsig_bddx.py.

Please pull the latest main and let me know if anything else comes up. Thanks again for the detailed report!