Dataset card

v1 relabelled: refined stuck rule, unittest counts (same records)

v1 + tau2 slice (train 3k, dev 1k, separate test_tau2); progress masked for tool domain

v1: all six questions, pairs, 2 prefixes per trajectory

v0 rebuilt: padded lines collapsed, think steps hidden, strict prefix filter

v0 p_success files (train 50k, dev 10k, test 64k)

initial commit