Live browser demo
Give your search a second check
A two-layer model scores the first 20 keyword matches.
Search the same 5,183 scientific abstracts used in the study. A pretrained TinyBERT reads each query–document pair and returns a relevance score, without writing a reply.
The model (about 17.6 MB) loads when you run the search.
What runs?
MS MARCO TinyBERT has two layers and about 4.39 million parameters. Both layers run for each selected pair, truncated to 256 tokens. The optional rule skips the entire reranker when the normalized keyword-score gap is at least 0.05. This is the study's frozen rule, not a calibrated reliability guarantee.
Results may support or refute a scientific claim; this is not a truth classifier. Your own query has no supplied relevance labels, so no accuracy percentage is invented. The model cannot recover a relevant paper missing from the 20 candidates.
Measured test results · Compare keyword and meaning search · All live demos