Reciprocal Rank Fusion (RRF) subsampling uses only 31.7% of documents while maintaining the full corpus's model ranking on Combined Recall@1000.
Reciprocal Rank Fusion (RRF) subsampling uses only 31.7% of documents while maintaining the full corpus's model ranking on Combined Recall@1000.
For pplx-embed-v1-4b, this reduces evaluation cost from 4,608 H200 GPU hours to about 1,500. https://t.co/ZPTvJ1xYoN