DataComp-VLM: Improved Open Datasets for Vision-Language Models

Hugging Face Daily Papers Papers

Summary

This paper introduces DataComp-VLM (DCVLM), a comprehensive benchmark for evaluating data curation strategies for vision-language models. The authors find that data mixing, rather than filtering, significantly improves performance, and their resulting DCVLM-Baseline dataset achieves state-of-the-art results on 33 downstream tasks.

Building performant Vision-Language Models (VLMs) requires carefully curating large-scale training datasets, yet the community lacks systematic benchmarks for evaluating such curation strategies. We introduce DataComp for VLMs (DCVLM), a benchmark for controlled data-centric experiments to improve VLM training. As part of DCVLM, we collect 160 datasets spanning four data types -- image-caption pairs, multimodal interleaved documents, text-only, and instruction-tuning data -- into a corpus of 6T multimodal tokens. DCVLM allows participants to test curation strategies (filtering, mixing, formatting, sampling) across 1B-8B models and 6.25B-200B token budgets. Models are then evaluated on a carefully selected suite of up to 52 downstream benchmarks across 9 domains. We conduct extensive experiments on DCVLM and find that data mixing, not filtering, is key to a high-quality training dataset: instruction-heavy mixtures scale better than caption-heavy ones, with gains widening at larger scales. The resulting dataset, DCVLM-Baseline, enables training an 8B VLM to 63.6% accuracy on our 33-task core suite with 200B training tokens. Compared to FineVision, the state-of-the-art open VLM training dataset, this represents an improvement of +5.4pp. DCVLM and all accompanying artifacts will be made publicly available at https://www.datacomp.ai/dcvlm/.
Original Article
View Cached Full Text

Cached at: 07/06/26, 10:36 AM

Paper page - DataComp-VLM: Improved Open Datasets for Vision-Language Models

Source: https://huggingface.co/papers/2606.28551 Authors:

,

,

,

,

,

,

,

,

,

,

,

,

,

,

,

,

,

,

,

,

Abstract

DataComp for VLMs (DCVLM) establishes a comprehensive benchmark for evaluating data curation strategies in vision-language models, demonstrating that data mixing rather than filtering significantly improves model performance at scale.

Building performantVision-Language Models(VLMs) requires carefully curating large-scale training datasets, yet the community lacks systematic benchmarks for evaluating such curation strategies. We introduce DataComp for VLMs (DCVLM), a benchmark for controlled data-centric experiments to improve VLM training. As part of DCVLM, we collect 160 datasets spanning four data types -- image-caption pairs, multimodal interleaved documents, text-only, and instruction-tuning data -- into a corpus of 6Tmultimodal tokens. DCVLM allows participants to test curation strategies (filtering, mixing, formatting, sampling) across 1B-8B models and 6.25B-200B token budgets. Models are then evaluated on a carefully selected suite of up to 52downstream benchmarksacross 9 domains. We conduct extensive experiments on DCVLM and find thatdata mixing, not filtering, is key to a high-quality training dataset: instruction-heavy mixtures scale better than caption-heavy ones, with gains widening at larger scales. The resulting dataset, DCVLM-Baseline, enables training an 8B VLM to 63.6% accuracy on our 33-task core suite with 200B training tokens. Compared to FineVision, the state-of-the-art open VLM training dataset, this represents an improvement of +5.4pp. DCVLM and all accompanying artifacts will be made publicly available at https://www.datacomp.ai/dcvlm/.

View arXiv pageView PDFProject pageGitHub23Add to collection

Get this paper in your agent:

hf papers read 2606\.28551

Don’t have the latest CLI?curl \-LsSf https://hf\.co/cli/install\.sh \| bash

Models citing this paper0

No model linking this paper

Cite arxiv.org/abs/2606.28551 in a model README.md to link it from this page.

Datasets citing this paper0

No dataset linking this paper

Cite arxiv.org/abs/2606.28551 in a dataset README.md to link it from this page.

Spaces citing this paper0

No Space linking this paper

Cite arxiv.org/abs/2606.28551 in a Space README.md to link it from this page.

Collections including this paper0

No Collection including this paper

Add this paper to acollectionto link it from this page.

Similar Articles

CaVe-VLM-CoT: An Interpretable Vision-Language Model Framework

arXiv cs.AI

CaVe-VLM-CoT is a modular reflection-based agentic-RAG framework for vision-language models that enforces evidence-grounded reasoning through a five-stage pipeline, achieving 87.1% accuracy on ScienceQA and proposing a suite of 23 metrics for evaluation.