MMFineReason Multimodal Reasoning Traces

## Overview MMFineReason is a large-scale multimodal reasoning dataset containing approximately 1.8M samples and 5.1B solution tokens, with detailed chain-of-thought reasoning annotations distilled from Qwen3-VL-235B-A22B-Thinking. The dataset is delivered in Parquet with image and text modalities and targets visual-question-answering, image-text-to-text, and text-generation training for multimodal reasoning models. Content focuses on mathematical reasoning, science, and STEM visual reasoning tasks in English. ## Schema - image — image (bytes/path) — input image associated with the problem - question / prompt — string — the multimodal question or task prompt - answer — string — final reference answer - reasoning / thinking — string — long-form chain-of-thought trace distilled from Qwen3-VL-235B-Thinking - solution — string — full solution text including reasoning and answer - source / subject — string — origin subset or topic tag (math, science, STEM, etc.) - (additional metadata columns may be present per the HF dataset page) ## Sources - HuggingFace: https://huggingface.co/datasets/OpenDataArena/MMFineReason-1.8M-Qwen3-VL-235B-Thinking — license: apache-2.0 - Reasoning traces distilled from Qwen3-VL-235B-A22B-Thinking (Alibaba) - Referenced paper: arXiv:2601.21821 ## Methodology Per the publisher, MMFineReason aggregates multimodal reasoning prompts (math, science, STEM with visual content) and uses the Qwen3-VL-235B-A22B-Thinking model to generate long-form reasoning traces and solutions for each item. The 1.8M samples / 5.1B solution token scale indicates large-scale distillation aimed at closing the multimodal reasoning gap through open data-centric methods. Data is provided as Parquet shards compatible with `datasets`, `dask`, and `polars`. ## Known gaps & limitations Reasoning traces are synthetic outputs of a single teacher model (Qwen3-VL-235B-Thinking), so they inherit that model's biases, errors, and stylistic patterns; correctness of CoT steps is not independently verified. Coverage is English-only and skewed to math/science/STEM — generalization to other multimodal domains is not guaranteed. The dataset has not been deduplicated against public multimodal evaluation benchmarks, so benchmark leakage is a real risk. Source does not exhaustively document gaps; buyers should validate empirically before using for benchmark-sensitive training. ## Intended use & out-of-scope - IS for: supervised fine-tuning and distillation of multimodal/VLM reasoning models, chain-of-thought training, RAG over reasoning exemplars, research on data-centric multimodal reasoning. - NOT for: training models intended to be evaluated on common multimodal math/science benchmarks without first running leakage checks; production decision-making in regulated domains; use cases requiring human-verified ground-truth reasoning. _Federated dataset: 198 parquet shards, 77.25 GB total. Queries and downloads stream through the DataBazaar API._ _PII signals: us_phone×8 present in the sample. Common in public datasets (papers, logs) but worth knowing before joining with private data._ Original supplier listing: MMFineReason 1.8M — Multimodal Reasoning Traces (Qwen3-VL-235B-Thinking) 1.8M multimodal reasoning samples with 5.1B solution tokens distilled from Qwen3-VL-235B-A22B-Thinking. Image+text inputs with detailed chain-of-thought annotations for math, science, and STEM visual reasoning. Apache-2.0.