AlibabaQwenintern/gap-p2-data

GAP Latent-Supervision Data

This is P2, an extended version of the training data for Fill the GAP: A Granular Alignment Paradigm for Visual Reasoning in Multimodal Large Language Models (NeurIPS 2026), used in follow-up experiments. The release has 45,702 rows: 41,996 in train and 3,706 in validation. Some training examples are deliberately repeated (see "Oversampling"), so the training split holds 34,408 distinct examples, and the release holds 38,114 distinct examples overall.

P2 extends the data used in the paper (AlibabaQwenintern/gap-data, "P1"): it adds 1,129 examples from the official We-Math 2.0 SFT set, each repeated 5 times, and repeats 3,072 existing training examples. The paper's results use P1, not this version.

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

Composition

data_source	Train rows	Distinct train examples	Validation rows	Rows with latent supervision (train + val)
monet_Visual_CoT	16,465	16,449	1,812	13,622
chart	7,334	6,187	671	6,818
WeMath2_SFT	5,645	1,129	0	4,150
GEOQA	4,857	3,760	401	1,477
MM_Math	2,415	1,988	246	855
monet_Zebra_CoT_visual_search	2,240	2,236	280	910
monet_Zebra_CoT_count	920	832	91	149
geo_3k	860	656	79	325
MathVision	471	393	44	277
monet_CogCoM	404	401	43	378
monet_ReFocus	360	358	38	283
monet_Zebra_CoT_geometry	25	19	1	21
Total	41,996	34,408	3,706	29,265

Oversampling

• The 1,129 We-Math 2.0 SFT examples each appear 5 times (oversample_copy = 1 to 5).
• Another 3,072 training rows, marked p2_oversampled: true, repeat existing examples, mostly from chart, GEOQA, MM_Math and geo_3k.
• No validation prompt appears in the training split.

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.
5. Oversampling for the P2 run, as described above.

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 rows (need_latent: false) have no image in the assistant turn.

Files

Path	Contents
train.jsonl	41,996 training rows (256 MB)
val.jsonl	3,706 validation rows (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
images/wemath2_sft_latent.tar	wemath2_sft_latent/	WeMath2_SFT

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-p2-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/
└── wemath2_sft_latent/

Most image paths in the JSONL files are absolute and start with /root/data/; a small number (1,343 of 74,193) 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 row.

Field	Type	Description
id	string	Unique row 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 row uses latent supervision
accuracy	float	Base-model accuracy over 8 samples (absent for WeMath2_SFT)
reinfer_response	string	Raw teacher response (absent for WeMath2_SFT)
reinfer_model	string	Teacher model, "qwen3-vl-235b" (absent for WeMath2_SFT)
reinfer_prompt_template	string	Prompt template used for the teacher (absent for WeMath2_SFT)
source_dataset	string	WeMath2_SFT only: "We-Math/We-Math2.0-SFT"
gold_answer, gold_solution	string	WeMath2_SFT only: answer and solution from the upstream dataset
oversample_source_id, oversample_copy	string, int	WeMath2_SFT only: original example ID and copy number (1 to 5)
p2_oversampled	bool	Present and true on the 3,072 repeated rows from other sources

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-p2-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 (all rows except WeMath2_SFT) 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.

WeMath2_SFT rows come from We-Math's official SFT training set (We-Math/We-Math2.0-SFT). None of its 1,129 distinct examples duplicates any of the 1,740 We-Math testmini questions (exact text, exact image, or perceptual-hash match).

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 (and source_dataset for We-Math 2.0 SFT).
• 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).
• WeMath2_SFT rows were not re-inferred by the teacher (their reinfer_* fields are absent); their question text and solutions come from We-Math 2.0 SFT and keep its CC BY-NC 4.0 license.

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
We-Math 2.0 SFT (We-Math/We-Math2.0-SFT)	WeMath2_SFT	CC BY-NC 4.0
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.