Teacher-forced token entropy for all 4,572 complete endpoints represented in the
Polaris trajectory-v2 PRM dataset (4,379 train and 193 validation trajectories).
Scores use the exact Qwen3.5-4B text tower that collected the trajectories.
Canonical forks are reasoning tokens whose raw full-vocabulary entropy is greater
than or equal to a threshold. The four global thresholds (q80, q90, q95, and
q99) are fitted… See the full description on the dataset page:
https://huggingface.co/datasets/asingh15/polaris-trajectory-v2-token-entropy.
