SwimBird-SFT-92K
收藏资源简介:
SwimBird-SFT-92K是由阿里巴巴集团·Accio团队构建的多模态推理微调数据集,旨在支持可切换推理模式的训练。该数据集包含92,300条样本,涵盖纯文本、纯视觉及视觉-文本交织三种推理模式,数据源自Zebra-CoT、ThinkMorph、MathCanvas和OpenMMReasoner等多样化基准。通过系统性筛选和分类策略,数据集确保了不同视觉依赖程度的任务覆盖,包括视觉搜索、空间导航、几何推理等复杂场景。其核心价值在于突破传统固定推理模式的局限,为动态自适应多模态推理提供训练基础,显著提升模型在视觉密集型任务和文本逻辑任务中的综合表现。
SwimBird-SFT-92K is a multimodal reasoning fine-tuning dataset constructed by the Accio Team of Alibaba Group, designed to support training with switchable reasoning modes. This dataset contains 92,300 samples, covering three reasoning modes: pure text, pure vision, and visual-text interleaved. Its data is sourced from diverse benchmarks including Zebra-CoT, ThinkMorph, MathCanvas, and OpenMMReasoner. Through systematic screening and classification strategies, the dataset ensures coverage of tasks with varying degrees of visual dependence, including complex scenarios such as visual search, spatial navigation, and geometric reasoning. Its core value lies in breaking through the limitations of traditional fixed reasoning modes, providing a training foundation for dynamically adaptive multimodal reasoning, and significantly improving the comprehensive performance of models in visually intensive tasks and textual logical reasoning tasks.

- 1SwimBird: Eliciting Switchable Reasoning Mode in Hybrid Autoregressive MLLMs华中科技大学; 阿里巴巴集团·Accio团队 · 2026年



