Medical retrieval model trained on RTX 3090 rivals full dataset runs

tomaarsen · x · 2026-08-26

Tomaarsen released the mLateOn-medical model and a detailed blog post. Scaling experiments revealed that training 100k pairs for 75 minutes on an RTX 3090 achieves within 1.2 NDCG@10 points of the full million-pair run. The model utilizes Multi-Vector Encoder and shows strong performance on medical retrieval benchmarks.

Related event: Fine-tuning multi-vector retrieval on a single RTX 3090 beats general-purpose baselines(12 posts)→

Original post →

More from Apps

Apps channel →