Open access to hourly-updating satellite composites and a 10 km, 15-minute precipitation product removes a major data barrier for teams in radar-sparse regions, where PARDIG's satellite-only training keeps accuracy high. Hydrology, solar power forecasting, wildfire monitoring and agri-weather startups can build on the 1999–2026 archive and live streams under CC BY 4.0. Compared with NASA IMERG-Final, the 30%+ gain in light-precipitation skill makes it a credible replacement baseline for downstream modelling work.
#open datasets
研究者给机器人 VLA 模型做了个潜层探针 VLA-Tug,能提前发现失败、还能预测失败原因,附带 3,180 条标注数据集。
Hugging Face 开了个新项目 ProtoStar,把公共资助的科学数据整理成 AI 能直接用的数据集,做科研或训练模型的朋友可以看看怎么参与。
BAAI 和人大做了个测 AI 读论文能力的基准,被 NeurIPS 2026 收了,数据和评测环境全开源,做科研智能体的可以直接拿去用。
Monsoon 是个 10 万小时的多语言语音数据集,微调 Whisper 后孟加拉语词错率从 85% 降到 7.65%,做语音方向的可以试试 API 测测效果。
This significantly lowers the barrier to entry for developing multilingual speech AI systems, particularly for underrepresented languages. Researchers and developers in ASEAN can leverage this comprehensive resource to build more inclusive speech technologies without the usual data collection challenges, potentially accelerating local language AI applications compared to proprietary alternatives.
Eidon AI 关门前把 1274 小时第一人称人类操作数据放上 Hugging Face,做具身智能的免费拿去用。
Harvey's budget research approach offers a blueprint for smaller organizations to compete with larger AI labs.
搞具身智能的团队,想给机器人加上手感,直接拿新智具身和复旦开源的3万小时触觉数据去训练,省时省力。
这批数据太香了,覆盖豆包、DeepSeek、Kimi等8个平台的搜索引用,21万条记录随便分析,搞AI搜索和信源生态的别错过。
Hugging Face 上悄悄放了 80TB 天文数据,用你笔记本就能跑 SDSS 和 Gaia 交叉匹配。搞 AI for Science 的话别错过,比想象中好上手。
GitHub新出的多语言数据集,免费开源,里面各种语言的README和讨论都有,做多语言AI模型训练正好用上。