Reinforcement Learning (GRPO/DAPO)

This type is specifically tailored for the reinforcement learning stage of LLMs (such as preference alignment algorithms like GRPO and DAPO). Centered on prompts and reward models (or evaluation criteria), the dataset drives models to achieve autonomous alignment and leaps in reasoning capabilities during policy iteration by providing explicit ground truths or evaluation rules.

Data format example

To make the field hierarchy clearer, the single-line JSONL data has been expanded.

{
  "prompt": [
    {
      "role": "user",
      "content": "Natalia sold paperclips to 48 of her friends in April, and she sold half as many in May as she did in April. How many paperclips did Natalia sell in total in April and May? Think step by step, and output the final answer after "####"."
    }
  ],
  "reward_model": {
    "ground_truth": "72"
  }
}

Field specifications

The table below shows an example of each entry (line) in the JSONL file.