BLIP3o Pretrain Long-Caption Dataset

## Overview BLIP3o-Pretrain-Long-Caption is a large-scale image-text dataset containing approximately 27 million images, each paired with a long descriptive caption (~120 tokens) generated by Qwen/Qwen2.5-VL-7B-Instruct. It is distributed in WebDataset (.tar) format suitable for streaming-based training of vision-language models. ## Schema - image — binary (JPEG/PNG) — the raw image - caption / txt — string — ~120-token long-form caption generated by Qwen2.5-VL-7B-Instruct - __key__ — string — WebDataset sample key/id ## Sources - BLIP3o team on Hugging Face — https://huggingface.co/datasets/BLIP3o/BLIP3o-Pretrain-Long-Caption — license: Apache-2.0 - Captions generated using Qwen/Qwen2.5-VL-7B-Instruct ## Methodology The BLIP3o team assembled a large image corpus and synthetically generated long-form descriptive captions for each image using the Qwen2.5-VL-7B-Instruct vision-language model. Captions average roughly 120 tokens, providing richer descriptions than typical alt-text or short-caption datasets. Data is sharded into WebDataset .tar archives for efficient streaming during distributed training. ## Known gaps & limitations Captions are model-generated and may contain hallucinations, factual errors, or systematic biases inherited from Qwen2.5-VL-7B-Instruct. The source image distribution and any filtering criteria are not fully documented on the dataset card. No demographic, geographic, or content-safety auditing is documented. Buyers should validate caption quality empirically and apply safety filtering before downstream use. ## Intended use & out-of-scope - IS for: pretraining and fine-tuning vision-language models, multimodal alignment research, caption-conditioned image generation training, dense captioning evaluation. - NOT for: tasks requiring human-verified ground-truth captions, safety-critical applications, or any use where caption factuality must be guaranteed. Not deduplicated against common multimodal eval suites — leakage risk for benchmark training. _Federated dataset: 10 parquet shards, 4.62 GB total. Queries and downloads stream through the DataBazaar API._ Original supplier listing: BLIP3o Pretrain Long-Caption Dataset (27M Images) 27 million images paired with ~120-token long captions generated by Qwen2.5-VL-7B-Instruct. WebDataset format, Apache 2.0. Ideal for vision-language pretraining and multimodal model fine-tuning.