OWL-SFT

# OWL SFT (Planner) Dataset ## Dataset Summary **OWL SFT** is a supervised fine‑tuning dataset designed for training the *planner* agent in the Optimized Workforce Learning (OWL) framework – a system for multi‑agent assistance in real‑world task automation. The dataset contains **1,564** multi‑turn conversations, focusing on **task decomposition, sequencing, and coordination** skills that are crucial for high‑level planning. ## Languages All conversation turns are written in **English**. ## Dataset Structure ### Data Fields | Column | Type | Description | | --------------- | ------------- | ------------------------------------------------------------------------------------------- | | `task_id` | *string* | Unique identifier for each task instance. | | `question` | *string* | Original user query or high‑level task goal. | | `conversations` | *list\[dict]* | Ordered list of dialogue turns, each with `role` (`user` / `assistant`) and `content` keys. | ### Data Splits Only a single **train** split is provided. Users may create validation/test splits via random sampling or k‑fold cross‑validation as required. ### Data Instances ```json { "task_id": "3b78e7c6", "question": "Plan a weekend trip to Kyoto for two people.", "conversations": [ {"role": "user", "content": "I want to spend a weekend in Kyoto with my partner."}, {"role": "assistant", "content": "Sure! Let me break this down into travel, accommodation, food, and activities..." } ] } ``` #### Personal & Sensitive Information The dataset does **not** contain personally identifiable information. All examples are synthetic or anonymised. ## Additional Information ### Citation ```bibtex @article{hu2025owl, title={Owl: Optimized workforce learning for general multi-agent assistance in real-world task automation}, author={Hu, Mengkang and Zhou, Yuhang and Fan, Wendong and Nie, Yuzhou and Xia, Bowei and Sun, Tao and Ye, Ziyu and Jin, Zhaoxuan and Li, Yingru and Chen, Qiguang and others}, journal={arXiv preprint arXiv:2505.23885}, year={2025} } ```