Idanwo ti o sopọ mọ itọsọna · Lile Ipele

Iṣeduro Iṣiṣẹ ati Aṣoju Idanwo Imọ-ẹrọ

Checks how activation steering vectors are estimated, applied at inference and scaled, along with risks from prompt-dependent generalization.

Jẹmọ itọsọna awọn ọnaIṣeduro Iṣeduro Ati Imọ-ẹrọ Oniduro
Ibeere 1 ti 8

Ni akoko wo ni iṣẹ awoṣe naa ṣe idari idari imuṣiṣẹ?

Awotẹlẹ gbogbo ibeere8 ibeere

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

  1. Ni akoko wo ni iṣẹ awoṣe naa ṣe idari idari imuṣiṣẹ?
    • Ni akoko itọkasi, nipa iyipada awọn iṣẹ ṣiṣe ti o farapamọ
    • Lakoko ikẹkọ iṣaaju nikan
    • Nigba data gbigba
    • Nikan lẹhin ti ọrọ ti o jade ti wa ni titẹ
  2. Bawo ni fekito idari ni igbagbogbo ṣe iṣiro?
    • Nipa ipilẹṣẹ matrix laileto
    • Bi awọn tumosi iyato laarin awọn ibere ise lori contrasting rere ati odi ta
    • Nipa piparẹ awọn fẹlẹfẹlẹ lati awoṣe
    • Nipa kika awọn ami-ami ni kiakia
  3. Iwoye wo ni o wa labẹ idi ti idari imuṣiṣẹ ṣiṣẹ?
    • Ero ti awọn awoṣe tọju data bi ọrọ itele
    • Igbagbọ pe gbogbo awọn neuronu jẹ aami kanna
    • Isọtẹlẹ oniduro laini, pe ọpọlọpọ awọn ẹya jẹ awọn itọnisọna laini aijọju
    • Nipe ti awọn awoṣe ko le soju fun awọn agbekale
  4. Kini olùsọdipúpọ irẹjẹ lori iṣakoso fekito idari?
    • Iwọn fokabulari awoṣe
    • Awọn nọmba ti fẹlẹfẹlẹ ninu awọn awoṣe
    • Oṣuwọn ikẹkọ ikẹkọ
    • Bawo ni ipa ihuwasi ti wa ni titari ni itọsọna ti o yan
  5. Kini ewu ti a mọ ti idari awoṣe kan ju ibinu lọ?
    • Iṣọkan ti iṣelọpọ ati awọn agbara le dinku
    • Awọn awoṣe reluwe ara laifọwọyi
    • Iwọn faili naa dagba lainidii
    • O n atunkọ awọn iwuwo patapata
  6. Ilana wo ni igbagbogbo so pọ pẹlu idari lati wa awọn ẹya itumọ lati dimole?
    • JPEG funmorawon
    • Awọn oluyipada autoencoders ti ko ṣoki ti o sọ awọn iṣẹ ṣiṣe di awọn ẹya
    • Iforukọsilẹ orukọ-ašẹ
    • Disiki defragmentation
  7. How does activation steering differ from fine-tuning the model weights?
    • It retrains the model on a new labeled dataset
    • It changes the tokenizer vocabulary
    • It adds a vector to hidden activations during inference without updating the base weights
    • It replaces the attention layers with a classifier
  8. A steering vector works on the prompts used to compute it but not on a different topic. Which limitation does this illustrate?
    • The vector must be an optimizer state
    • Steering always requires fine-tuning
    • A residual stream contains no features
    • The direction found from one prompt set may not generalize to other contexts