Papers like BioMed-R1 discuss how many examples in training are knowledge-heavy and not reasoning-heavy and filtering out such examples can help training.
Write a ligthweight script that, given a HuggingFace dataset like https://huggingface.co/datasets/open-thoughts/OpenThoughts2-1M or https://huggingface.co/datasets/GeneralReasoning/GeneralThought-430K, filters/tags only samples that require reasoning and not just knowledge.
Papers like BioMed-R1 discuss how many examples in training are knowledge-heavy and not reasoning-heavy and filtering out such examples can help training.
Write a ligthweight script that, given a HuggingFace dataset like https://huggingface.co/datasets/open-thoughts/OpenThoughts2-1M or https://huggingface.co/datasets/GeneralReasoning/GeneralThought-430K, filters/tags only samples that require reasoning and not just knowledge.