linagora/synthetic_corrections

合成修正数据集包含300个建筑师与建造者之间的对话，这些对话是基于Minecraft结构化对话语料库的互动启发的。每个对话样本包含三条建筑师指令，建造者需要根据这些指令在网格上构建简单的形状，其中两次构建正确，一次构建错误。建筑师会对错误的构建进行纠正。数据集的目标是训练一个能够预测响应建筑师纠正动作序列的代理模型。每个样本还包含了对话回合之间的语义关系。

The Synthetic Corrections Data corpus consists of 300 dialogues between Architect and Builder, inspired by the interactions found in the Minecraft Structured Dialogue Corpus. Each dialogue sample contains three instructions from the Architect, which require the Builder to construct simple shapes on a grid, with two of the constructions being correct and one being incorrect. The Architect provides a Correction for the incorrect construction. The task for an agent model acting as the Builder is to predict the action sequence in response to the Architects Correction. Each sample also includes the semantic relationships between dialogue turns represented as typed tuples in the Structure field.