The Extrapolation Cliff in On-Policy Distillation of Near-Deterministic Structured Outputs
Paper • 2605.08737 • Published • 2
ErrOPD
When does on-policy distillation transfer teacher errors? Paper, checkpoints and data. Code: github.com/LiXin97/ErrOPD
Note The paper (arXiv 2605.08737).
XINLI1997/ErrOPD-ESCI
Viewer • Updated • 29k • 37Note ESCI listwise ranking, clean and error-injected teacher SFT lists (Apache-2.0).
XINLI1997/ErrOPD-Spider
Viewer • Updated • 4.09k • 30Note Spider text-to-SQL pools (CC BY-SA 4.0).
XINLI1997/ErrOPD-Fashion
Viewer • Updated • 37.9k • 42Note Amazon Fashion review ranking with Gemini 2.5 Pro scores (Amazon Reviews 2023 derived, research use only).
-
XINLI1997/ErrOPD-Qwen3.5-0.8B-student-Fashion
Text Generation • 0.9B • Updated • 56 -
XINLI1997/ErrOPD-Qwen3.5-0.8B-Fashion-earlyclose-rKL
Text Generation • 1B • Updated • 53 -
XINLI1997/ErrOPD-Qwen3.5-0.8B-Fashion-earlyclose-fKL
Text Generation • 1B • Updated • 52 -
XINLI1997/ErrOPD-Qwen3.5-9B-teacher-earlyclose-Fashion
Text Generation • 9B • Updated • 56 -
XINLI1997/ErrOPD-Qwen3.5-4B-Fashion-earlyclose-fKL
Text Generation • 5B • Updated • 61 -
XINLI1997/ErrOPD-Qwen3.5-4B-Fashion-earlyclose-rKL
Text Generation • 5B • Updated • 58 • 1 -
XINLI1997/Llama-3.2-3B-ErrOPD-Fashion-dupid-rKL
Text Generation • 4B • Updated • 76 -
XINLI1997/ErrOPD-Qwen3.5-2B-Fashion-earlyclose-fKL
Text Generation • 3B • Updated • 50 -
XINLI1997/ErrOPD-Qwen3-1.7B-ESCI-earlyclose-fKL
Text Generation • 2B • Updated • 53 -
XINLI1997/ErrOPD-Qwen3-1.7B-ESCI-dupid-rKL
Text Generation • 2B • Updated • 54 • 1 -
XINLI1997/ErrOPD-Qwen3-1.7B-student-SQL-start3pct
Text Generation • 2B • Updated • 54 -
XINLI1997/ErrOPD-Qwen3-1.7B-student-SQL
Text Generation • 2B • Updated • 51 -
XINLI1997/Llama-3.2-3B-ErrOPD-Fashion-earlyclose-rKL
Text Generation • 4B • Updated • 77 -
XINLI1997/Llama-3.2-3B-ErrOPD-student-Fashion
Text Generation • 175k • Updated • 79 -
XINLI1997/ErrOPD-Qwen3.5-2B-Fashion-earlyclose-rKL
Text Generation • 3B • Updated • 54 -
XINLI1997/ErrOPD-Qwen3-1.7B-ESCI-earlyclose-rKL
Text Generation • 2B • Updated • 53 -
XINLI1997/ErrOPD-Qwen3-1.7B-SQL-fKL
Text Generation • 2B • Updated • 54 • 1 -
XINLI1997/ErrOPD-Qwen3-1.7B-student-ESCI
Text Generation • 2B • Updated • 53 -
XINLI1997/Llama-3.2-3B-ErrOPD-Fashion-earlyclose-fKL
Text Generation • 4B • Updated • 77 • 1 -
XINLI1997/Llama-3.2-3B-ErrOPD-Fashion-dupid-fKL
Text Generation • 4B • Updated • 69 -
XINLI1997/ErrOPD-Qwen3.5-2B-student-Fashion
Text Generation • 2B • Updated • 53 -
XINLI1997/ErrOPD-Qwen3-1.7B-ESCI-earlyclose-rKL-lambda1.40
Text Generation • 2B • Updated • 53 -
XINLI1997/ErrOPD-Qwen3-1.7B-ESCI-dupid-fKL
Text Generation • 2B • Updated • 50 • 1 -
XINLI1997/Llama-3.1-8B-ErrOPD-teacher-earlyclose-Fashion
Text Generation • 266k • Updated • 75 • 1 -
XINLI1997/Llama-3.1-8B-ErrOPD-teacher-dupid-Fashion
Text Generation • 266k • Updated • 77 • 1 -
XINLI1997/ErrOPD-Qwen3-1.7B-SQL-rKL
Text Generation • 2B • Updated • 93 -
XINLI1997/ErrOPD-Olmo-3-7B-student-Fashion
Text Generation • 528k • Updated • 93 -
XINLI1997/ErrOPD-Olmo-3-7B-Fashion-earlyclose-rKL
Text Generation • 7B • Updated • 90 -
XINLI1997/ErrOPD-Olmo-3-7B-Fashion-earlyclose-fKL
Text Generation • 7B • Updated • 94 -
XINLI1997/ErrOPD-Qwen3.5-4B-student-Fashion
Text Generation • 5B • Updated • 90 -
XINLI1997/ErrOPD-Qwen3-4B-teacher-dupid-ESCI
Text Generation • 196k • Updated • 164 -
XINLI1997/ErrOPD-Olmo-3.1-32B-teacher-earlyclose-Fashion
Text Generation • 1.05M • Updated • 127 -
XINLI1997/ErrOPD-Qwen3-4B-teacher-SQL
Text Generation • 196k • Updated • 125 -
XINLI1997/ErrOPD-Qwen3-4B-teacher-earlyclose-ESCI
Text Generation • 196k • Updated • 149