Act-PRM inferred-thought generations for snorkel_finance agent traces
(Qwen3-4B-Instruct-2507). For each logged (state s, action x), an offline EM
samples G=4 candidate thoughts z, scores each by the length-penalized action
likelihood, and commits the top-1.
One row per logged action step. Full candidate pool so you can take top-1 OR recompute
any weighting:… See the full description on the dataset page:
https://huggingface.co/datasets/mzio/aprm-thought-generations-snorkel-finance.
