opendatalab/ChartVerse-RL-40K

ChartVerse-RL-40K是一个专门为强化学习设计的图表推理数据集，包含最具挑战性的样本。这些样本具有最高的失败率，即最困难的样本，即使强大的视觉语言模型也难以解决，但仍有可能解决。数据集包含40K个样本，每个样本都有唯一的图表，且经过验证的答案准确性。数据集的生成过程包括三个步骤：1) 使用Rollout Posterior Entropy (RPE)确保图表的高结构复杂性；2) 使用Truth-Anchored Inverse QA Synthesis生成QA对；3) 选择最难样本，即失败率最高的样本。数据集适用于视觉问答、图像文本到文本和强化学习等任务。

ChartVerse-RL-40K is a curated dataset of the most challenging chart reasoning samples for Reinforcement Learning. This dataset contains samples with the highest failure rates — the most difficult samples that strong VLMs struggle with but can still solve occasionally. These samples provide the strongest learning signal for RL training. The dataset includes 40K samples, each with unique charts and verified answer accuracy. The data generation pipeline involves three steps: 1) Ensuring high structural complexity with Rollout Posterior Entropy (RPE); 2) Generating QA pairs using Truth-Anchored Inverse QA Synthesis; 3) Selecting the hardest samples with the highest failure rates. The dataset is suitable for tasks such as visual-question-answering, image-text-to-text, and reinforcement learning.