Data filtering based on reasoning difficulty

[tmabraham]

Write a lightweight script that, given a HuggingFace dataset like https://huggingface.co/datasets/open-thoughts/OpenThoughts2-1M or https://huggingface.co/datasets/GeneralReasoning/GeneralThought-430K, and a baseline model, and reward function, filters/tags the questions with their reward.

[khankanz]

I can take a stab at this over the weekend.

[khankanz]

If a separate 'reference_answer' doesn't exist for a given ds i.e. OpenThoughts2. Should we opt for using the 'model_answer' field? If so, entire CoT or look for a format like \\(\\boxed{181}\\). for the final answer?