VillanovaAI/Multi_PixMoCap-QA

Multi-PixMo-Cap-QA

Overview

Multi-PixMo-Cap-QA is a multilingual version of Multi-PixMo-Cap dataset by Villanova. Both datasets are based on the PixMo series developed by AllenAI.

The original PixMo-Cap dataset was created by recording annotators speaking freely about an image for 60–90 seconds, then transforming the resulting audio transcripts into detailed captions using Claude (see the PixMo paper).

Multi-PixMo-Cap followed the same multimodal concept, but all examples were re-generated from human captions using a permissively licensed model: Qwen3-VL-235B-A22B-Instruct. This results in a multilingual dataset that mirrors the PixMo-Cap style while being fully based on open-source generation methods.

Multi-PixMoCap-QA dataset used the same images and captions of the Multi-PixMo-Cap dataset and used them to generate a Q&A set for each sample. The model was provided with the image, the caption and a predefined list of question types. After viewing the image, the model selected the most suitable question type from the list and generated a corresponding Q&A pair.

Model:

The dataset was generated using the multimodal, open LLM Qwen/Qwen3.5-122B-A10B-FP8.

Language Distribution

Multi-PixMo-Cap-QA contains five subsets, one per language, distributed as follows:

Language	Subset Name	Percentage	Number of Examples
English	en	32%	103737
Italian	it	17%	55486
French	fr	17%	55486
Spanish	es	17%	55486
German	de	17%	55486

English accounts for 32% of all examples, while the other four languages collectively make up 68%, distributed evenly.

Question type distribution

Since the model autonomously selected the most fitting task type for each image, the distribution of question types across the dataset is uneven. Below is a table showing the percentage distribution for each task.

Total dataset entries: 325681

Question type	Count	Percentage
text_recognition	76076	23.36%
attribute_and_properties	70747	21.72%
formatting_task	53797	16.52%
aesthetic_judgement	41040	12.60%
creative_writing	29637	9.10%
emotion_interpretation	29410	9.03%
actions_and_events	14386	4.42%
comparative_reasoning	3930	1.21%
count_measure	1889	0.58%
identify_object	1646	0.51%
scene_and_context	1633	0.50%
common-sense_world-knowledge	1486	0.46%
categorization_tagging	4	0.00%

Intended Use

This dataset aims to provide multilingual question-answer pairs suitable for training multimodal models.

It also offers a Q&A corpus produced by a permissively licensed model to ensure that the resulting dataset can be freely reused, including for the development and training of new models.

Dataset Structure

Each subset contains the following fields:

Field	Type
image_url	string
question_type	string
question	string
answer	string

Licensing

The dataset is distributed under the ODC-BY-1.0 license. Users must adhere to the attribution and usage requirements outlined under this license.