Idanwo ti o sopọ mọ itọsọna · Alabọde Ipele

LLM-bi-a-adajọ adanwo

Review model-based evaluation, pointwise and pairwise rubrics, judge biases and validation against human preferences. in practice.

Jẹmọ itọsọna awọn ọnaLlm bi onidajọ
Ibeere 1 ti 8

Kini 'LLM-bi-adajọ' tọka si?

Awotẹlẹ gbogbo ibeere8 ibeere

Awọn idahun ati awọn alaye han bi o ṣe n mu ibeere naa.

  1. Kini 'LLM-bi-adajọ' tọka si?
    • Lilo awoṣe ede kan lati ṣe Dimegilio tabi ṣe afiwe awọn abajade awoṣe miiran
    • Lilo awoṣe ede lati kọ awọn idajọ ofin
    • Awoṣe ti o dahun bẹẹni/ko si ibeere nikan
    • Awoṣe ti o ṣe ipo awọn abajade wiwa nipasẹ ọjọ
  2. Kini iyato laarin pointwise ati pairwise idajo?
    • Pointwise ni yiyara; pairwise jẹ losokepupo sugbon aami bibẹẹkọ
    • Pointwise onipò ọkan idahun lodi si a rubric; pairwise mu awọn dara ti meji
    • Pointwise nikan ṣiṣẹ fun koodu; pairwise nikan fun prose
    • Ko si iyato gidi
  3. Ewo ninu iwọnyi jẹ aibikita ti o ni akọsilẹ daradara ni awọn onidajọ LLM?
    • Nfẹ awọn idahun ti a kọ sinu gbogbo awọn lẹta nla
    • Nigbagbogbo fẹran ọrọ akọkọ ni adibi
    • Favoring gun idahun laiwo ti didara
    • Kiko lati awọn ibeere ite nipa itan
  4. Bawo ni awọn oluyẹwo ṣe maa n koju abosi ipo ni awọn afiwera meji?
    • Nipa fifi idahun kan han ni akoko kan
    • Nipa fifi idahun to gun sii ni akọkọ
    • Nipa piparẹ aaye idalare
    • Nipa nṣiṣẹ kọọkan lafiwe lemeji pẹlu awọn ibere swapped
  5. What evaluation procedure did G-Eval use for generated text?
    • Chain-of-thought steps followed by a form-filling evaluation against criteria
    • A speaker verification threshold
    • A search index with BM25 only
    • An image classifier trained on screenshots
  6. Bawo ni ẹgbẹ kan ṣe le jẹrisi awọn adajọ LLM kan ni ibamu pẹlu awọn ayanfẹ eniyan?
    • Nipa ko ṣayẹwo lẹẹkansi lẹhin iṣeto
    • Nipa calibrating ti o lodi si a eda eniyan-ike ṣeto goolu
    • Nipa ṣiṣe awọn onidajọ ite ara
    • Nipa lilo nikan ni awọn ipari ose
  7. What is self-preference bias in LLM-based evaluation?
    • A judge always selects the shortest answer
    • A judge cannot compare two outputs
    • A judge favors outputs generated by its own model or model family
    • A judge evaluates only audio signals
  8. What does G-Eval use to make its discrete score more fine-grained in the reported method?
    • The position of the answer in a pairwise comparison
    • The judge model’s parameter count only
    • A count of citations in the answer
    • Probability weights over the candidate score tokens