Check out our open source repo for large scale multimodal training examples with LanceDB!
3 examples published so far:
→ Object detection (AV perception): BDD100K + Feature Engineering + Faster R-CNN
→ ViT: MFU benchmark across LanceDB, S3, and Parquet backends
→ VLA: Ray +
A dataset dump gets you raw files, embedding and indexing still yours to build.
EuroSAT in Lance format on @huggingface ships with OpenCLIP embeddings and an IVF_PQ index already built. 27,000 tiles, search and class filters work on load.
Browse the dataset:
7 days for model training down to 1. 95% GPU utilization. 100K+ QPS.
@ByteDanceOSS's Volcano Engine rebuilt its AI stack on Lance and LanceDB, from training data to ArkClaw agent memory.
Read more:
LeRobot's X-VLA soft-fold dataset in Lance format: 1,542 episodes, three camera streams, 2.85M frames.
Every MP4 segment stored inline with timestamp bounds, so you can pull exactly the clip you need without decoding the rest of the video.
Schema, blob access, training code:
