pipeline_tag: text-classification
library_name: transformers
Boundary Predictor - TESSY
This model is a boundary predictor developed as part of the work: How to Fine-Tune a Reasoning Model? A Teacher–Student Cooperation Framework to Synthesize Student-Consistent SFT Data.
Model Description
This model serves as a critical component of TESSY (Teacher–Student Cooperation Data Synthesis), a framework designed to synthesize "on-policy" SFT data for reasoning models like Qwen3.
The primary role of this predictor is to identify reasoning boundaries—the specific points in a sequence where the model transitions between reasoning steps (capability tokens) and non-reasoning stylistic content (style tokens). By predicting these boundaries, TESSY allows a teacher model (to provide reasoning) and a student model (to provide stylistic consistency) to alternate during the data generation process.
Resources
- Paper: How to Fine-Tune a Reasoning Model? A Teacher-Student Cooperation Framework to Synthesize Student-Consistent SFT Data
- GitHub Repository: CoopReason/TESSY
- Associated Dataset: TESSY-Code-80K
Citation
If you find this work or model useful, please cite:
@article{TESSY,
title={How to Fine-Tune a Reasoning Model? A Teacher--Student Cooperation Framework to Synthesize Student-Consistent SFT Data},
author={Huang, Zixian and Yang, Kaichen and Huang, Xu and Hao, Feiyang and Qiming, Ge and Li, Bowen and Du, He and Chen, Kai and Guo, Qipeng},
journal={arXiv preprint arXiv:2604.14164},
year={2026}
}