Prueba de ajuste de modelos de incrustación para recuperación

Las respuestas y explicaciones aparecen a medida que realiza el cuestionario.

1. What is a hard negative in embedding fine-tuning?
   • A passage that looks relevant to the query but does not answer it
   • A passage from a different language
   • Any randomly chosen passage
   • A query with no answer
2. Why are hard negatives more useful than easy negatives?
   • They are cheaper to store
   • The base model already separates easy negatives, so hard ones teach new distinctions
   • They remove the need for positives
   • They shorten training time to zero
3. What is a false negative in this context?
   • A positive pair with a typo
   • A mined negative that actually does answer the query
   • A query that retrieves nothing
   • A vector with all zeros
4. Which loss is commonly used for training embedding models on query-passage pairs?
   • Mean squared error on token counts
   • A contrastive loss such as InfoNCE (MultipleNegativesRankingLoss)
   • Cross-entropy on next-word prediction only
   • Hinge loss on image pixels
5. After switching to a newly fine-tuned embedding model, what must you do to your document index?
   • Nothing; old vectors still work
   • Re-embed the entire collection with the new model
   • Delete half the documents
   • Convert vectors to text
6. Which alternative might give larger gains than fine-tuning with less effort?
   • Removing all metadata
   • Improving chunking, adding BM25 hybrid search or a reranker
   • Using a smaller context window
   • Lowering the embedding dimension to 2
7. Which metrics does the guide recommend for evaluating retrieval?
   • BLEU and ROUGE
   • recall@k and nDCG@10
   • Perplexity only
   • Training loss only
8. Why should synthetic questions generated from documents be filtered?
   • They are always in the wrong language
   • They often copy the document's wording, making the task too easy
   • They cannot be embedded
   • They violate the loss function