JonesLin/openmmreasoner-sft-concise-cot-314k

OpenMMReasoner-SFT concise CoT (314K)

Concise, human-readable rewrites of the reasoning traces in OpenMMReasoner/OpenMMReasoner-SFT-874K: one trace per question, 314,471 questions. The rewritten traces keep every value, equation and intermediate result the reasoning uses and drop restatements, filler, hedging and repeated checks. In total they have 14.2M words, 7.8% of the 182.6M words of the traces they replace (about 45 words per trace instead of about 580).

The records are text only. Each one points at the OpenMMReasoner row and image its question comes from, so the images can be taken from that dataset.

Fields

field
id	our id for the question
omr_config	the OpenMMReasoner-SFT-874K config the question comes from: mmr1, m1_sft, llava_cot, OpenVLThinker-sft-iter3 or WeMath
omr_id	the id of the source row in that config
omr_image	the image path in OpenMMReasoner-SFT-874K's image_zip archives; empty for text-only questions (m1_sft)
question	the question text, as in the source row
cot	the concise reasoning trace
answer	the final answer, as in the source row
cot_source	rewritten (main pass), repaired (second pass, see below) or original (the source trace, kept)

omr_config	questions		cot_source	questions
mmr1	169,661		rewritten	308,978
m1_sft	85,287		repaired	4,624
llava_cot	55,921		original	869
OpenVLThinker-sft-iter3	2,913
WeMath	689

How it was made

1. Questions. The source traces were deduplicated and limited to 8,192 tokens (Qwen3.5 tokenizer). Questions were removed when their image is a near duplicate (10-bit dHash) of an image in, or their text shares a 13-gram with, one of these evaluation sets: ChartQA, MathVista, MMBench, AI2D, HallusionBench, MMStar, MathVision, MathVerse, We-Math, DynaMath and MMMU. One trace was kept per remaining question.

2. Rewriting. gpt-6-luna (reasoning effort low) rewrote the traces in batches of 20 (traces of 300 words or more) or 40 (shorter ones). For each item it saw the question text, the original trace and the answer, not the image. The instructions were:

   Rewrite each reasoning trace below to be concrete and concise, while staying easy for a person to read.

   • Concrete: keep every specific value, quantity, equation and intermediate result the reasoning uses, including what is read from the image.
   • Concise: drop restating the question, filler, hedging and repetition.
   • Readable: write in plain, complete sentences that a reader can follow step by step; use equations inline where they help.
   • Keep the same reasoning path and reach the same final answer. Do not add facts or numbers that are not in the original.

3. Repair. Rewrites that switched language, lost LaTeX backslashes, kept only the answer, came back longer than the original or were missing were redone at reasoning effort medium, with two more rules: write in the same language as the original trace, and escape every backslash. 869 traces still failed these checks and kept their original text (cot_source = original).

Known limitations

• LaTeX is often rewritten as Unicode text (15π for 15\pi, 90° for 90^\circ), so the notation of cot can differ from the notation of answer.
• In a check of 300 held-out questions rewritten the same way, about 1% of the rewrites (proof questions) say that the original argument does not establish the claim instead of concluding it, so they disagree with answer.
• The rewriter did not see the images: what a trace says about an image is only what the original trace said.
• About 2–3% of the rewrites contain a number that does not appear in the original trace. In spot checks these were mostly correct intermediate values or artifacts of the check.

Terms

The questions, answers, images and original traces come from OpenMMReasoner-SFT-874K and the datasets it collects; the rewritten traces were generated with gpt-6-luna. Use of this dataset is subject to the terms of those sources and of the generating model.