Molmo2 Open Training Data

Molmo2 is Ai2's (the Allen Institute for AI) open vision-language model family, released in December 2025, and this entry covers the training data published alongside it rather than the model weights. The Molmo2 Data collection on the Hugging Face Hub gathers 13 datasets built for video and multi-image understanding: dense video captioning (Molmo2-Cap), video question answering (VideoCapQA, VideoSubtitleQA, AskModelAnything), spatio-temporal pointing, counting and point-based multi-object tracking (VideoPoint, VideoCount, VideoTrack), and multi-image QA and pointing (MultiImageQA, SynMultiImageQA, MultiImagePoint), several with matched evaluation splits. You access them freely through the Hub as Parquet, loadable with the datasets library, Dask or Polars. The source videos are not bundled: the files hold annotations and reference YouTube video IDs, so the clips must be fetched separately using the download scripts in the allenai/molmo2 code repository. For RAG and AI work the collection is primarily supervised training and grounding data for multimodal models rather than a retrieval index, and it is strongest for teams building or fine-tuning open VLMs that read video, track objects or reason across several images, the encoders and instruction followers that sit inside a multimodal RAG pipeline. The licence position needs care. The data is offered under ODC-BY, which requires attribution but not share-alike, and Ai2 frames it for research and educational use under its Responsible Use Guidelines. Two caveats matter for commercial use: some captions were generated with proprietary models (GPT-4.1 and GPT-5), so OpenAI's terms attach, and some subsets derive from third-party academic datasets limited to non-commercial research, documented in a separate source-attribution file. Commercial use therefore requires per-dataset checking rather than relying on the ODC-BY headline. The technical report (arXiv 2601.10611) lists the training sources. It sits alongside the open multimodal training corpora we catalogue such as FineVision, but is focused on video and multi-image grounding rather than general image-text instruction tuning.