BLIP3o-60k
收藏资源简介:
BLIP3o-60k是一个高质量指令调整数据集,用于图像生成。该数据集包含60,000条数据,由GPT-4o使用覆盖各种场景、对象、人类手势等的多样化描述进行提示而生成。数据集的创建旨在提高视觉审美和指令遵循能力,通过在BLIP3o-60k上进行监督指令调整,显著提高了BLIP3-o与人类偏好的一致性,并改善了美学质量。该数据集可用于训练和评估统一的跨模态模型,支持图像理解和生成任务。
BLIP3o-60k is a high-quality instruction tuning dataset for image generation. Comprising 60,000 instances, it was created by prompting GPT-4o with diverse descriptions covering various scenarios, objects, human gestures, and other relevant content. The dataset is designed to enhance visual aesthetics and instruction-following capabilities of models. Supervised instruction tuning on BLIP3o-60k has significantly improved the alignment between BLIP3-o and human preferences, as well as the aesthetic quality of generated content. This dataset can be used to train and evaluate unified cross-modal models, supporting both image understanding and image generation tasks.
BLIP3o-60k 数据集概述
基本信息
- 语言: 英文 (en)
- 数据集名称: BLIP3o-60k
- 数据规模: 10K<n<100K
数据集内容
数据集包含以下类别的文本到图像指令调优数据,这些数据从GPT-4o中提取:
- JourneyDB
- Human (包括带有MSCOCO人类标注、人类手势、职业的数据)
- Dalle3
- Geneval (与测试集无重叠)
- Common objects
- Simple text
下载与使用
下载代码
python from huggingface_hub import snapshot_download snapshot_download(repo_id=BLIP3o/BLIP3o-60k, repo_type=dataset)
读取数据集代码
python from datasets import load_dataset import glob data_files = glob.glob(/your/datasets/path/*.tar) train_dataset = load_dataset("webdataset", data_files=data_files, cache_dir=/your/cache/directory/, split="train", num_proc=64)

- 1BLIP3-o: A Family of Fully Open Unified Multimodal Models-Architecture, Training and DatasetSalesforce Research · 2025年



