DAPO-17.4k

This dataset contains the 17,398-row DAPO math set in the RL parquet format used by verl.

Each prompt asks the model to return the final answer in \boxed{}:

Please output the final answer within \boxed{}.

Splits

• train: 16,398 examples
• validation: 1,000 examples

The validation set was sampled uniformly at random from dapo-math-17k.parquet with seed 42; the train split contains the remaining examples. The splits have no overlap by extra_info.index.

Schema

• data_source: source label, currently math
• prompt: single-turn user prompt with the boxed-answer instruction
• reward_model: rule reward metadata with ground_truth
• extra_info: original question, stable index, and split label