Kediska Models-ka

Jawaabaha iyo sharraxaadaha waxay soo baxayaan markaad qaadanayso kediska.

1. Tuubada dib-u-soo-celinta caadiga ah ee laba-heer ah, waa maxay shaqada dib-u-dejiyaha?
   • In dib loo dhaliyo oo dib loo habeeyo liiska musharixiinta si ay u khusayso
   • In laga soo saaro musharaxiinta bilowga ah malaayiin dukumeenti ah
   • Si loo kaydiyo dukumeentiyada kaydka xogta
   • Si loo dhaliyo jawaabta ugu dambeysa ee luqadda dabiiciga ah
2. Besides MS MARCO, on which dataset did Nogueira and Cho report strong results for their BERT passage reranker?
   • ImageNet
   • TREC-CAR
   • Hadal-qoraal
   • SQuAD only
3. Waa maxay sababta aan cod-bixiye-ku-beddele loo isticmaali karin si toos ah malaayiin dukumeenti ah loo baadho?
   • Waxay kaliya ku shaqeysaa sawirada
   • Ma akhrin karo qoraal
   • Dhibcaheeda waa in loo xisaabiyaa mid cusub lammaane kasta oo dukumeenti ah, kaasoo aad u gaabis ah marka loo eego cabbirka
   • Wax xisaabin ah uma baahna
4. Waa maxay faa'iidada ugu weyn ee labada-encoder ka sarreeya cod-geliyaha?
   • Way ka saxsan tahay marka la eego khusaysa-garan
   • Had iyo jeer waxay dib u dalbataa natiijooyinka si qumman
   • Uma baahna xog tababar
   • Ku dhejinta dukumeentigeeda waa la sii xisaabin karaa, taasoo ka dhigaysa soo-saarid degdeg ah oo la cabbiri karo
5. How can a reranker affect context quality in a retrieval-augmented generation pipeline?
   • It can move more relevant retrieved candidates higher in the context shortlist, but does not guarantee correctness
   • It ensures every retrieved passage is factually correct
   • It creates missing documents in the corpus
   • It removes the need to evaluate retrieval
6. In listwise reranking, what does the model consider together?
   • Only one document with no query
   • A query and a set of candidate documents when deciding their ordering
   • Each document embedding before any candidate retrieval
   • A single relevance label for the full training corpus
7. What should a team verify before interpreting reranker scores as probabilities?
   • Whether every score is above zero
   • Whether the first-stage retriever was removed
   • Whether the scores have been calibrated and evaluated for that use
   • Whether the model has more parameters than the retriever
8. What limitation does a reranker inherit from the first-stage retrieval system?
   • It always returns a calibrated probability
   • It removes the need for retrieval
   • It guarantees a correct generated answer
   • It cannot promote a relevant document that was never included in the candidate set