CrossImage-CoT-36K

36,495 multi-image questions, each with a structured chain-of-thought rationale. The rationale format is Understand → Perceive (one line per image) → Connect (relate the images) → Conclude, followed by <answer>…</answer>.

This is the SFT data behind Kaylee0501/Qwen3-VL-8B-CrossImage-SFT.

How the rationales were made

1. No gold answer was shown to the teacher. Each question went to one GPT teacher (gpt-5.6-luna, gpt-5.4, gpt-5.6-terra, gpt-4.1, gpt-4o or o4-mini). The teacher did not see the answer.
   • A trace was kept only if the teacher's first answer matched the gold answer. Wrong answers were never resampled, and no item was retried with another model.
   • The reason: showing the gold makes teachers invent perceptions that fit the answer. In our audit, 31% of gold-shown traces on items the teacher could not solve unaided misdescribed an image.
2. Format check. The Connect step must cite at least two images. A text-only judge checked that open-ended rationales support the gold answer.
3. Answer stored as the gold. The final <answer> holds the gold, which equals the teacher's answer for auto-verified items.
4. Option letters rebalanced. For each number of options, the counts of gold letters were balanced to remove letter shortcuts.
5. Decontaminated. Every sample with an image near-identical to an image in BLINK, MuirBench, MIRB or Mantis-Eval was removed (dHash ≤ 4, then mean pixel difference < 6 at 32×32): 2,788 samples in total.
6. Synthetic families with programmatic gold answers:
   • synthetic_jigsaw: BLINK-style fill-the-hole puzzles
   • synthetic_mirb_jigsaw: MIRB-style 2×2 piece ordering
   • synthetic_correspondence: a photo plus the same photo after a known perspective warp and relighting, with four labelled candidate points
   • abstention pairs, where "None of the choices provided" is the correct answer

Fields

field	meaning
uid	sample id
category, source, task_family	task type and origin
images	image paths, in order, as Image 1, Image 2, …
question	the question with options and the CoT instruction, as used for training
answer	the rationale plus <answer>gold</answer>, the training target
gold	gold answer (an option letter, an image number or text)
teacher_model	the model that wrote the rationale
visual_audit	pass / fail / null. A blinded audit by a model that did not write the trace checked each "Image k" line against the image. fail means at least one per-image claim was judged clearly false. Auditor precision on a manual check was about 85%.
second_teacher_agrees	whether a second, independent teacher also answered correctly without the gold (null = not checked: open-ended or audit-failed rows)

Filtering to visual_audit == "pass" gives 30,987 rows. Adding second_teacher_agrees != false on top of that gives 28,564 rows.

Images

• synthetic/… paths: the 27,404 synthetic images are in images/synthetic.tar. Extract it so these paths resolve.
• All other paths: they point into lmms-lab/M4-Instruct-Data using its own relative layout, for example imagecode/0.0.0/<hash>/train_images/.... Download and unzip those archives into the same root.

Composition

category	rows	category	rows
Retrieval	6,016	Temporal	2,733
Difference	4,847	Multi-view	2,671
Jigsaw	4,340	Reasoning	2,113
Similarity	4,135	Consistency	1,744
Abstention	3,409	Counting	881
Correspondence	2,852	Ordering	754

teacher	rows
gpt-5.6-luna	9,252
gpt-5.4	7,708
gpt-5.6-terra	7,333
gpt-4.1	6,009
gpt-4o	5,328
o4-mini	865

There are 37 sources in total. The M4 sources include imagecode, contrastive_caption, nuScenes, nextqa, STAR, RecipeQA, MagicBrush, NLVR2, IconQA, VISION, Spot-the-Diff, Birds-to-Words, TQA, WebQA, CLEVR-Change, MIT-States, IEdit, AESOP, multi_vqa and ScanQA.

License and intended use

• Research use only. The images and questions come from the M4-Instruct source datasets and keep their licenses. Some are non-commercial; for example, nuScenes is CC BY-NC-SA 4.0.
• Synthetic images are crops, warps and composites of those source images and inherit the same terms.
• Rationales were generated with OpenAI models through Azure. Use them in line with the provider's terms.