Beat the baseline & ship·5 min
Recall@k, refreshed
0 / 1
Time to measure the model honestly. recall@k = the fraction of queries whose gold document lands in the top-k results by cosine. For our frozen eval set, each query's gold is its known paraphrase; the rest of the pool are distractors.
recall@5 over 4 queries
query 1: gold at rank 2 ✓ (in top 5)
query 2: gold at rank 1 ✓
query 3: gold at rank 9 ✗ (outside top 5)
query 4: gold at rank 4 ✓
recall@5 = 3 / 4 = 0.75
Deterministic ties
When two candidates have equal cosine, break ties by index (ascending) so the ranking is reproducible. Otherwise float ties could shuffle the top-k between runs and make recall non-deterministic.