4papersubmission/fmts

FMTS (Failure-Mined Transition Stress) 是一个用于跟踪转换压缩的合成对话基准数据集。每个对话包含一个已知的转换（old_state, op, new_state, support_span）和一个依赖于new_state的探针。数据集包含多个配置，如fmts_explicit（显式话语标记）、fmts_implicit（无表面标记）、sgd_scope（SGD重新切割为仅用户的中对话转换）和multiwoz_scope（MultiWOZ重新切割为相同协议）。数据集主要用于诊断目的，不用于训练。数据集是合成的，由模板生成，操作分布偏向REPLACE_VALUE。数据集遵循CC-BY 4.0许可证。

FMTS (Failure-Mined Transition Stress) is a synthetic dialogue benchmark for transition-tracking compression. Each dialogue has one known transition (old_state, op, new_state, support_span) and one probe whose answer depends on new_state. The dataset includes multiple configurations such as fmts_explicit (explicit discourse markers), fmts_implicit (no surface markers), sgd_scope (SGD recut to USER-only mid-dialogue transitions), and multiwoz_scope (MultiWOZ recut to the same protocol). The dataset is intended for diagnostic purposes only and not for training. It is synthetic, template-generated, and the operation distribution is skewed towards REPLACE_VALUE. The dataset is licensed under CC-BY 4.0.