Finetuned ColBERT Model Performs on Medical Tasks
Posts highlight finetuned ColBERT performance on medical tasks and storage reductions via quantization.
Omar Khattab retweeted a post praising mxbai-edge-32M results from a blog post on training ColBERT models with Sentence Transformers for medical retrieval. Florian Brand replied that new results on hierarchical pooling and asymmetric quantization make storage excuses for late interaction models harder to sustain. A separate retweet noted asymmetric quantization as a free lunch. One generated summary described storage for a medical model dropping from 45 GB to 2.8 GB with NDCG@10 falling only from 0.916 to 0.
Combined views
422
3 posts, first seen 1d ago
Finetuned ColBERT Model Performs on Medical Tasks
Posts highlight finetuned ColBERT performance on medical tasks and storage reductions via quantization.
Omar Khattab retweeted a post praising mxbai-edge-32M results from a blog post on training ColBERT models with Sentence Transformers for medical retrieval. Florian Brand replied that new results on hierarchical pooling and asymmetric quantization make storage excuses for late interaction models harder to sustain. A separate retweet noted asymmetric quantization as a free lunch. One generated summary described storage for a medical model dropping from 45 GB to 2.8 GB with NDCG@10 falling only from 0.916 to 0.
Reactions
3 postsRT 
Rikiya Takehi@rikiyatakehiVery cool! Also happy to see how our tiny mxbai-edge-32M holds up:)
- likes: 30
- replies: 0
- bookmarks: 8
- reposts: 1
AA Aamir@aaxsh18y’all should use asymmetric quantization. its literally one of the few free lunches
- likes: 33
- replies: 2
- bookmarks: 19
- reposts: 3
FB Florian Brand@xeophon@lateinteraction becoming harder and harder to find bad excuses
- likes: 9
- replies: 1
- bookmarks: 0
- reposts: 0