delinqu/

comet-1.5k

This repository releases Comet-1.5k, an RFT (Rejection Sampling Fine-Tuning) dataset generated by Team Comet’s post-training pipeline on the BEHAVIOR-1K simulator (OmniGibson).It is intended to support reproducible research on post-training for VLA policies, dataset construction via rejection sampling, and efficient iteration over high-signal rollouts.

Context. Team Comet’s BEHAVIOR-1K challenge submission achieved Q-score 0.2514 on Held-out… See the full description on the dataset page: https://huggingface.co/datasets/delinqu/comet-1.5k.