OnceSelect: Reusable Data Selection for Efficient Multimodal Instruction Tuning

Name: OnceSelect: Reusable Data Selection for Efficient Multimodal Instruction Tuning
Uploaded: 2026-10-09T05:34:56.513638+00:00
Duration: 608 s
Description: The OnceSelect framework solves inefficient multimodal instruction tuning by creating a reusable data selector trained once. It uses frozen CLIP features to cluster data and train a lightweight selector to identify samples with low confidence scores, retaining diverse and informative examples from any new dataset efficiently.

The OnceSelect framework solves inefficient multimodal instruction tuning by creating a reusable data selector trained once. It uses frozen CLIP features to cluster data and train a lightweight selector to identify samples with low confidence scores, retaining diverse and informative examples from any new dataset efficiently.

Open the full episode