OpenDataArena/MMFineReason-SFT-586K-Qwen3-VL-235B-Thinking

MMFineReason-SFT-586K 是 MMFineReason-1.8M 数据集的一个子集，包含了 Qwen3-VL-4B-Thinking 模型无法一致成功的 33% 最难样本（通过率 ≠ 1）。该数据集专注于需要非平凡多模态推理的挑战性样本。主要特点包括 586K 个挑战性样本、平衡的难度覆盖以及来自 Qwen3-VL-235B-A22B-Thinking 的高质量思维链（CoT）注释。数据集还展示了性能结果，表明在更难样本上训练可以提高模型性能。MMFineReason-SFT-586K is a subset of the MMFineReason-1.8M dataset, containing the hardest 33% of samples where the Qwen3-VL-4B-Thinking model does not consistently succeed (pass rate ≠ 1). This dataset focuses on challenging samples that require non-trivial multimodal reasoning. Key highlights include 586K challenging samples, balanced difficulty coverage, and high-quality Chain-of-Thought (CoT) annotations from Qwen3-VL-235B-A22B-Thinking. The dataset also features performance results showing that training on harder samples improves model performance.