TESSY-Code-80K
- Type
- dataset
- Venue
- CoopReason
- Year
- 2026
- Source
- huggingface
- Access
- free
- Language
- English, Python (and other programming languages)
- Added
- 2026-07-17T20:18:03.802256+00:00
- Verified
- 2026-07-17T20:18:03.802256+00:00
Summary
TESSY-Code-80K is an 80k-row programming contest training dataset synthesized using the TESSY (Teacher-Student Cooperative Data Synthesis) framework, where a teacher model (GPT-OSS-120B) generates reasoning content and a student model (Qwen3-8B) generates stylistic content. This cooperative approach produces on-policy SFT data that preserves teacher reasoning quality while maintaining student distribution consistency, improving Qwen3-8B by up to 11.34% on LiveCodeBench-Pro and 6.68% on OJBench.
Keywords
code reasoning teacher-student on-policy sft synthetic programming-contest qwen3 gpt-oss
Topics
Code
Research notes
- Associated with arXiv:2604.14164 and GitHub repo CoopReason/TESSY. Source problems collected from OpenThoughts and NVIDIA Nemotron datasets. Teacher-only data causes performance drops; TESSY cooperative synthesis avoids this.