DataComp-VLM

DataComp-VLM is a competition to curate the best training dataset to (pre)train Vision-Language Models. We release a standardized pool of up to 6T multimodal tokens, and participants can experiment with filtering, mixing, and formatting samples to create the best data. The resulting models are evaluated on a suite of upto 52 multimodal tasks.

📦 Releases

Candidate pools

Here we provide links to the uncurated (raw) pools that each competition scale filters and mixes from. Pools are nested: each smaller pool is a subset of the next.

Scale	Pool size	🤗 Hub
small	187.5B tokens	link
medium	750B tokens	link
large	3T tokens	link
x-large	6T tokens	link

Pre-mixed, ready-to-train datasets

Here we provide read-to-train fully decontaminated, globally shuffled, and pre-mixed data mixtures — no curation or mixing needed, point any training stack at the shards and go! Mixture is given as captioning / instruction / text / multimodal-docs, by training samples.

Dataset	Scale	Mixture	Tokens	Samples	Shards	Size	🤗 Hub
DCVLM-Baseline (6.25B)	small	10 / 70 / 15 / 5	6.25B	3,253,356	384	0.77 TB	link
DCVLM-Baseline (200B)	x-large	10 / 70 / 15 / 5	200B	103,985,276	10,752	25.7 TB	link
DCVLM-Balanced (200B)	x-large	40 / 40 / 15 / 5	200B	112,358,849	11,318	19.7 TB	link

Models

Model	Scale	Trained on	Vision init.	LLM init.	🤗 Hub
DCVLM-1B	small	DCVLM-Baseline	InternViT-300M	Qwen2.5-0.5B	link
DCVLM-2B	medium	DCVLM-Baseline	InternViT-300M	Qwen2.5-1.5B	link
DCVLM-4B	large	DCVLM-Baseline	InternViT-300M	Qwen2.5-3B	link
DCVLM-8B	x-large	DCVLM-Baseline	InternViT-300M	Qwen2.5-7B	link

NOTE: We are still releasing all our artefacts gradually at the links above. If a dataset / model has not yet been updated at their respective link, they are currently queued and will be updated as soon as possible. We appreciate your patience as we work through all our uploads! We are also cleaning up our training, eval and data infra code, we will release those soon too!