paper
TL;DR#
- I read this because.. : github 돌아다니다가 curriculum이라서
- task : reasoning LLM
- problem : curriculum 학습 하고 싶다 (deepscaleR과 비슷)
- idea : prompt가 길면 더 복잡할 것이다
- architecture :
- objective : GRPO loss
- baseline : DEEPSEEK-R1DISTILL-QWEN-1.5B, STILL-1.5B7, DeepScaleR1.5B-Preview, RSTAR-MATH-7B , QWEN-2.5-MATH-7B-Instruct, QWEN2.5-7B-SimpleRL8, and EURUS-27B-PRIM
- data : AIME problems (1984-2023), AMC problems (before 2023), Omni-MATH dataset, Still dataset
- evaluation : MATH 500, AIME 2024, AMC 2023, Minerva Math, and OlympiadBench
- result : baseline대비 좋은 성능. training cost가 deepscaleR의 50%
- contribution :
- etc. :
Details#
