uiuc-convai/CALM-IT

CALM-IT是一个多任务数据集，旨在通过整合多轮对话管理（TOD）和函数调用能力（LA）来训练对话代理语言模型（CALM）。它结合了ReAct-based推理和复杂的API使用，能够实现面向任务的对 robust 对话（TOD）和动态工具交互。CALM-IT被用于训练CALM模型系列（8B、70B、405B），在三个基准测试中超过了GPT-4o：MultiWOZ 2.4（TOD）、BFCL V3（LA）和API-Bank（LA）。

CALM-IT is a multi-task dataset designed to train Conversational Agentic Language Models (CALM) by integrating multi-turn dialogue management (TOD) with function-calling capabilities (LA). It interleaves ReAct-based reasoning with complex API usage, enabling robust task-oriented dialogue (TOD) and dynamic tool interaction. CALM-IT was used to train the CALM model series (8B, 70B, 405B), outperforming GPT-4o across three benchmarks: MultiWOZ 2.4 (TOD), BFCL V3 (LA), and API-Bank (LA).