Users praise the LightOn-Rerank models release for its honest discussion of unsuccessful experiments and strong results on mixed text-image RAG corpora.
Based on 2 visible X reactions from 3 accounts; directional sample.
Ask a question below.
Published answers will appear here.
The blog is very honest about what didn't work, which is the best part: - Tournament scheduling: -27 NDCG - First-token readout: capped at (low) 1.6x gain - Reading a listwise model pointwise: throws away most of the lift - For 2B, the ViT encoder was the slowest, not the decode
Strong release for teams doing RAG over mixed corpora (PDFs, page scans, plain text). Definitely give these a try! Great work to @coreprinciple_ and @AmelieTabatta.
Users praise the LightOn-Rerank models release for its honest discussion of unsuccessful experiments and strong results on mixed text-image RAG corpora.
Based on 2 visible X reactions from 3 accounts; directional sample.
Ask a question below.
Published answers will appear here.