MMFineReason-SFT Multimodal Reasoning Samples

## Overview MMFineReason-SFT-123K is a difficulty-filtered multimodal reasoning dataset containing ~123,000 visual-question-answering samples paired with chain-of-thought reasoning traces from Qwen3-VL-235B-Thinking. It is the hardest 7% slice of the larger MMFineReason-1.8M corpus — specifically, samples where Qwen3-VL-4B-Thinking failed all 4 inference attempts (pass rate = 0). The dataset spans mathematics, science, and STEM visual reasoning tasks, distributed in parquet with image+text modalities. ## Schema - `image` — image — the visual context for the reasoning problem - `question` — string — the natural-language prompt/question - `answer` — string — ground-truth final answer - `thinking` / `reasoning` — string — Qwen3-VL-235B-Thinking chain-of-thought trace - `source` — string — original benchmark/source of the sample - `category` — string — topic area (math, science, STEM subdomain) - `difficulty` — indicator — confirmed hard (0/4 pass rate by 4B thinking model) - additional metadata columns for provenance ## Sources - HuggingFace: https://huggingface.co/datasets/OpenDataArena/MMFineReason-SFT-123K-Qwen3-VL-235B-Thinking — Apache-2.0 - Parent corpus: MMFineReason-1.8M (OpenDataArena) - Reasoning traces generated by Qwen3-VL-235B-Thinking ## Methodology The publisher started from MMFineReason-1.8M, a large multimodal reasoning corpus aggregated from STEM/math/science visual QA sources. They ran Qwen3-VL-4B-Thinking 4 times per sample and retained only items with a 0/4 pass rate, yielding ~123K "hardest 7%" examples. For each retained sample, Qwen3-VL-235B-Thinking was used to produce a long-form reasoning trace plus final answer, forming SFT-ready (image, question, thinking, answer) tuples. The publisher reports comparable downstream performance to training on the full 1.8M with only 7% of the data. ## Known gaps & limitations Difficulty filtering is defined relative to a single 4B reference model, so "hardness" is model-specific and may not transfer to all student models. Reasoning traces are model-generated by Qwen3-VL-235B and may contain hallucinated steps even when final answers are correct. Coverage is English-only and skewed toward math/science/STEM visual problems — not representative of general multimodal tasks. Source-level deduplication against public eval suites (MathVista, MMMU, etc.) is not documented; benchmark leakage risk should be assessed before use in evaluation training. Source does not document additional gaps; buyers should validate empirically. ## Intended use & out-of-scope - IS for: SFT / distillation of multimodal reasoning into smaller VLMs, hard-sample fine-tuning, chain-of-thought training for visual STEM tasks, ablation studies on data efficiency. - NOT for: benchmark evaluation (likely overlaps with public eval sets), non-STEM multimodal tasks, non-English use cases, or treating model-generated reasoning traces as ground truth. _Federated dataset: 18 parquet shards, 6.70 GB total. Queries and downloads stream through the DataBazaar API._ Original supplier listing: MMFineReason-SFT-123K (Qwen3-VL-235B Thinking Traces) 123K hardest-7% multimodal reasoning samples with Qwen3-VL-235B chain-of-thought traces. Curated SFT subset of MMFineReason-1.8M where smaller models consistently fail. Apache-2.0, parquet, image+text.