Jarrodbarnes/tau2-sft-seed-v3

Tau2 Seed SFT Dataset v3是一个用于在RFT（拒绝采样微调）之前进行SFT（监督微调）热启动的种子数据集。它具有以下关键特征：正确的双控制行为、无错误的工具调用和8个专家编写的电信黄金示例。数据集包含397个样本，分布在零售、航空和电信三个领域，其中电信领域有53个样本，包括5个完全成功的样本、40个高部分成功的样本和8个手工制作的黄金示例。黄金示例涵盖了飞机模式故障排除、SIM卡重新插入、APN重置、移动数据切换、速度测试+数据节省、漫游启用、网络偏好和转接人工等模式。数据集的使用流程包括SFT种子对齐策略、RFT生成样本、SFT种子+RFT成功样本和RL二进制奖励。

Tau2 Seed SFT Dataset v3 is a seed dataset for SFT warm-start before RFT (Rejection Sampling Fine-Tuning). It features correct dual-control behavior, no wrong tool calls, and 8 expert-written telecom golden examples. The dataset contains 397 samples across retail, airline, and telecom domains, with 53 samples in telecom, including 5 full successes, 40 high-partial successes, and 8 hand-crafted golden examples. Golden example patterns include airplane mode troubleshooting, SIM card reseat, APN reset, mobile data toggle, speed test + data saver, roaming enable, network preference, and transfer to human. The pipeline involves SFT on seed for policy alignment, RFT for generating samples, SFT on seed + RFT successes, and RL with binary reward.