R1-Zero-like training for math

Created on April 3|Last edited on April 27

Comment

Context: https://huggingface.co/spaces/open-r1/README/discussions/20﻿
Big-Math-RL-Verified experimentsExperiments run on a preprocessed version of the SynthLabsAI/Big-Math-RL-Verifieddataset.
Configs for training runs: https://github.com/huggingface/open-r1/pull/569﻿
Phase 1: stabilise training
Phase 2: tune for performance
DAPO Math 17k experimentsExperiments run on a processed version of the BytedTsinghua-SIA/DAPO-Math-17k dataset
Configs for training runs: https://github.com/huggingface/open-r1/pull/569﻿
﻿
 
Dr GRPO4
 
Overlong filtering2
 
Num iterations4
 
Prompt batch size3
 
Num tokens2
 
Ref model2
Fast generation5
﻿
﻿

Add a comment