AlibabaQwenintern/gap-data

GAP Latent-Supervision Data

This is the training data used for all results in Fill the GAP: A Granular Alignment Paradigm for Visual Reasoning in Multimodal Large Language Models (NeurIPS 2026). It contains 36,985 examples: 33,279 in train and 3,706 in validation, with no duplicates and no overlap between the splits.

The construction pipeline described in the paper produced 49,309 candidate examples. After preprocessing, 36,985 remain, and these are split 90/10 into the training and validation sets released here.

Each example pairs a query image and question with a structured teacher response. Examples that use latent supervision also mark where a visual latent goes and name the auxiliary image that supervises it.

An extended version used in follow-up experiments, which adds the official We-Math 2.0 SFT set and repeats some training examples, is available at AlibabaQwenintern/gap-p2-data. The paper's results use the data in this repository.

Composition

data_source	Train	Validation	Examples with latent supervision (train + val)
monet_Visual_CoT	16,449	1,812	13,608
chart	6,187	671	5,831
GEOQA	3,760	401	1,174
monet_Zebra_CoT_visual_search	2,236	280	910
MM_Math	1,988	246	730
monet_Zebra_CoT_count	832	91	134
geo_3k	656	79	253
monet_CogCoM	401	43	376
MathVision	393	44	233
monet_ReFocus	358	38	282
monet_Zebra_CoT_geometry	19	1	17
Total	33,279	3,706	23,548

How the data was built

1. Auxiliary image. For each training query, Nano Banana 2 (Gemini image generation) produces an intermediate auxiliary image intended to expose the visual evidence the question needs.
2. Structured response. The query image, question, and auxiliary image are given to Qwen3-VL-235B-A22B-Thinking (reinfer_model: qwen3-vl-235b), which writes the reasoning trace, a <parser> description of the intended auxiliary signal, and the final answer.
3. Quality filtering. Examples with degenerate auxiliary-image embeddings, or with <parser> outputs that fail a length or coherence check, are removed.
4. Difficulty-aware assignment. The base model (Qwen2.5-VL-7B) is sampled 8 times per query, and accuracy stores the fraction it solved. Only examples it never solved (accuracy 0) are marked need_latent: true. Examples it solved at least once keep text-only supervision.

Response format

Assistant turns follow this structure:

<think> reasoning ... <|vision_start|><|image_pad|><|vision_end|>
<parser> description of the auxiliary visual signal </parser> continued reasoning </think>
<answer> final answer </answer>

The image placeholder inside the assistant turn marks where the visual latent goes. The image it refers to (listed in that turn's image_filenames with image_types: "target") is the auxiliary image whose frozen-ViT features supervise the latent. In the paper's training code this placeholder becomes <|latent_start|>, 36 <|latent_pad|> tokens, and <|latent_end|>. Text-only examples (need_latent: false) have no image in the assistant turn.

Files

Path	Contents
train.jsonl	33,279 training examples (216 MB)
val.jsonl	3,706 validation examples (24 MB)
vit_pca_mu.npy	Mean vector of the vision-embedding PCA basis, shape (3584,), float32
vit_pca_W.npy	Projection matrix of the vision-embedding PCA basis, shape (3584, 629), float32: 629 components retaining 95% of the variance of 3584-d vision embeddings
images/*.tar	Image archives, one per source folder (below); about 50 GB in total

Archive	Extracted folder	data_source values
images/monet.tar	monet/	monet_Visual_CoT, monet_Zebra_CoT_*, monet_CogCoM, monet_ReFocus
images/chart.tar	chart/	chart
images/GEOQA.tar	GEOQA/	GEOQA
images/MM_MATH.tar	MM_MATH/	MM_Math
images/MATHVISION.tar	MATHVISION/	MathVision
images/geo3k_train.tar	geo3k_train/	geo_3k

Download and restore

Access is gated: log in to Hugging Face and accept the conditions on this page before downloading.

hf auth login
hf download AlibabaQwenintern/gap-data --repo-type dataset --local-dir /root/data
cd /root/data
for f in images/*.tar; do tar -xf "$f" -C /root/data; done

With older versions of huggingface_hub, run huggingface-cli download with the same arguments.

After extraction the layout is:

/root/data/
├── train.jsonl
├── val.jsonl
├── vit_pca_mu.npy
├── vit_pca_W.npy
├── monet/
├── GEOQA/
├── MATHVISION/
├── MM_MATH/
├── chart/
└── geo3k_train/

Most image paths in the JSONL files are absolute and start with /root/data/; a small number are relative to that same root. Keep the data at /root/data, or rewrite the /root/data/ prefix and resolve relative paths against your own root.

Data fields

Each line of train.jsonl and val.jsonl is one example.

Field	Type	Description
id	string	Unique example identifier (UUID)
data_source	string	Source category (see Composition)
data_type	string	Always "interleave"
message	list	Two turns, user then assistant (fields below)
reference_answer	string	Reference answer text
need_latent	bool	Whether the example uses latent supervision
accuracy	float	Base-model accuracy over 8 samples
reinfer_response	string	Raw teacher response
reinfer_model	string	Teacher model, "qwen3-vl-235b"
reinfer_prompt_template	string	Prompt template used for the teacher

Fields of each message turn:

Field	Description
role	"user" or "assistant"
dialog	Turn text; images appear as <|vision_start|><|image_pad|><|vision_end|>
image_filenames	Image paths for the placeholders in dialog, in order
image_types	"input" for query images, "target" for auxiliary latent targets (omitted when the turn has no image)
need_mask	Always false in this release

Usage

After hf auth login and accepting the access conditions:

from datasets import load_dataset

ds = load_dataset("AlibabaQwenintern/gap-data")  # train / validation metadata from the JSONL files
print(ds)

This loads the JSONL metadata only. Images come from the extracted archives (see "Download and restore").

Leakage audit

We checked the curated corpus against the evaluation benchmarks used in the paper (HRBench4K, MMStar, MME-RealWorld-Lite, MathVista, We-Math, POPE). There are no exact image duplicates and no exact normalized question-text duplicates. Some high-similarity geometry cases remain, because GEOQA and Geometry3K are also reused by public evaluation benchmarks, so we do not claim the training and evaluation distributions are fully disjoint.

Licenses

This dataset does not have a single license. Each example combines content taken from an upstream dataset with content we added, and each part keeps its own license:

• Upstream content (the query image and question text taken from a source dataset) keeps the license and terms of that source, listed below. The source of every example is recorded in data_source.
• Our additions (auxiliary images, teacher responses and reasoning traces, latent-supervision markup, the need_latent and accuracy labels, and the PCA basis files vit_pca_mu.npy and vit_pca_W.npy) are released under CC BY-NC 4.0, for non-commercial research. The auxiliary images were generated with Nano Banana 2 through the Gemini API and are also subject to the Gemini API terms and prohibited-use policy; the teacher responses were generated with Qwen3-VL-235B-A22B-Thinking (Apache-2.0).

Using an example means following the terms of each of its parts.

Upstream source	data_source	Upstream license / terms
Visual-CoT (deepcs233/Visual-CoT), via Monet-SFT-125K (MIT)	monet_Visual_CoT	HF metadata lists Apache-2.0, but the dataset card states CC BY-NC 4.0 and research-purpose use; treated as CC BY-NC 4.0
Zebra-CoT (multimodal-reasoning-lab/Zebra-CoT), via Monet-SFT-125K (MIT)	monet_Zebra_CoT_visual_search, monet_Zebra_CoT_count, monet_Zebra_CoT_geometry	CC BY-NC 4.0
CogCoM / CoMDataset (qijimrc/CoMDataset), via Monet-SFT-125K (MIT)	monet_CogCoM	CC BY-SA 4.0 for dataset contributions; original image copyrights retained
ReFocus (ReFocus/ReFocus_Data), via Monet-SFT-125K (MIT)	monet_ReFocus	Apache-2.0
ChartQA (ahmed-masry/ChartQA; vis-nlp/ChartQA)	chart	GPL-3.0. The ChartQA authors crawled the chart images from online sources such as Statista and Pew Research Center; image copyrights remain with the original owners
MM-MATH (THU-KEG/MM_Math)	MM_Math	MIT. Problems are collected from Chinese middle-school exams
MathVision (MathLLMs/MathVision)	MathVision	MIT
GeoQA (hz2475/geoQA)	GEOQA	MIT
Geometry3K (xyliu6/geometry3k)	geo_3k	Apache-2.0

Intended use and limitations

• Intended for research on multimodal reasoning and visual latent modeling. Not intended for commercial use.
• Responses and auxiliary images are model-generated and may contain errors; the <parser> description states the intended auxiliary signal and is not a verified ground-truth caption.
• The data reflects the coverage of its source pools. In particular, real-world scenes are under-represented, which matches the weaker MME-RealWorld-Lite results reported in the paper.

Citation

@inproceedings{miao2026gap,
  title     = {Fill the {GAP}: A Granular Alignment Paradigm for Visual Reasoning in Multimodal Large Language Models},
  author    = {Miao, Yanting and Sun, Yutao and Wang, Dexin and Zhou, Mengyu and Poupart, Pascal and Lv, Lei and Zhao, Qi and Wang, Li and Li, Hao and Jiang, Xiaoxi and Jiang, Guanjun},
  booktitle = {Advances in Neural Information Processing Systems},
  year      = {2026}
}

Corresponding author: Mengyu Zhou.