WRIT-2K is a 2,000-trajectory supervised fine-tuning dataset for multi-turn, tool-using customer-service agents on tau2-bench style airline and retail tasks.
This dataset accompanies the paper WRIT: Write-Read Intensive Trajectory Synthesis for Multi-Turn User-Facing Agents.
Project homepage:
https://hengrui-gu.github.io/WRIT/WRIT-2K contains complete multi-turn trajectories with user messages, assistant natural-language responses… See the full description on the dataset page:
https://huggingface.co/datasets/Henryoung/WRIT-2K.
