Multi-domain SFT dataset for training tool-use agents on tau2-bench dual-control environments. Designed for use with the slime RL framework.
Metric
Value
Total trajectories
416
Domains
airline, retail, telecom
Format
+ [ACTION]
Train-only
Yes
Domain
Train Tasks
Coverage
airline
30
100%
retail
74
100%
telecom
74
82.4%
SFT Baseline (Qwen3-4B, 1 epoch)… See the full description on the dataset page: https://huggingface.co/datasets/Jarrodbarnes/tau2-sft-final.