Ikibazo gihuza ibibazo · Biroroshye Urwego
Ubumenyi bwo guca amatariki muri LLMs Ikibazo
Gerageza icyo kugabanya ubumenyi aribyo, kuki abanyamideli batamenya neza itariki yabo yo guhagarika, nuburyo gushakisha, kugarura no gutera inshinge bikomeza ibisubizo byubu.
Ikibazo 1 Bya 8
Itariki ntango y'ubumenyi y'icyitegererezo ni iki?
Reba buri kibazo8 ibibazo
Ibisubizo nibisobanuro bigaragara mugihe ufata ikibazo.
- Itariki ntango y'ubumenyi y'icyitegererezo ni iki?
- The date the model stops working
- The approximate date after which its training data contains little or no information
- The date the model was released to the public
- The date the vendor ends customer support
- Kuki ibyitegererezo bikunze kutizera, cyangwa se bigasuzugura, itariki yabyo ntango?
- Vendors deliberately hide it from the model
- Models delete their most recent data after training
- Models count days incorrectly
- The final months before the cutoff are thinly represented, because people keep writing about events long afterward
- Ni ubuhe buryo bwizewe cyane bwo kubona itariki ntango y'ubumenyi y'icyitegererezo?
- Check the vendor's published model documentation
- Ask the model directly
- Count the words in its answers
- Look at the model's logo
- Ni ubuhe buryo bworoshye bukosora ibibazo byinshi by'icyitegererezo cyibwira itariki mbi y'ubu?
- Fine-tuning the model every week
- Turning the model off and on
- Putting the current date in the system prompt
- Using shorter prompts
- Akarobot k'ikiganiro gafite ubushakashatsi bwa web gasubiza ikibazo cy'amakuru y'icyumweru gashize. Ni iyihe mvugo nyayo?
- Its training cutoff has moved to last week
- It can no longer make factual errors
- It retrains itself on every query
- Its training cutoff is unchanged; it is reading new sources and is only as accurate as the pages it retrieves
- Why is retrieval usually preferred over fine-tuning for keeping facts current?
- Fine-tuning is slower and more costly as a way to keep facts current
- Fine-tuning is impossible for large models
- Retrieval permanently changes the model's weights
- Retrieval works without any documents
- Which of these is a typical sign of stale knowledge?
- The model refuses every question
- The model switches to another language
- The model names an outdated software version as the latest
- The model's answers are unusually short
- When building evals for cutoff-related problems, what should you include?
- Only questions about events well before the cutoff
- Questions whose correct answers changed after the cutoff
- Questions with all dates removed
- A question asking the model to state its cutoff