179 questions
No questions match those filters.
← All questions
Retrieval & EmbeddingsHard· Cross-lingual Retrieval· Language Bias· Contrastive Loss
You shipped the multilingual encoder. Recall@100 barely...
This is one of the questions in the full AI/ML interview bank. Pro unlocks all 1789 questions; Premium includes the same bank plus the highest daily Practice limit.
See plansThe retrieval score is being dominated by a language-specific constant (the norm of the per-language mean) rather than semantic similarity. Because the model is trained with uniform in-batch negatives, matching the language of the query is an easy, ‘free’ way to minimize contrastive loss. Fine-tuning often exacerbates this by further rewarding the model for this shortcut. The fix requires language-matched negatives to remove the language signal as a discriminator.