LLM hallucinate a lokacin da ya samar da rubutu wanda ke bayyana da karfin yake, amma ba daidai ba ne ko ya yi ni aiki. Don farahinta, dole ne ka fahimta abin da modelar take yi a gaske.
A wurin da aka nuna, LLM shine . Bayan rubutu da ya kasance, yana hasashen wadda wadda waje mai yuwuwa (kalma/ratsin kashi) dangane da — ba ta hanyar binciken gida ba.
input: "The capital of Australia is"
model: P(next token) → "Canberra" 0.71, "Sydney" 0.18, ...
→ samples a token, appends it, repeats
LLMs suna kasua a cikin aiki wanda gaske masu amfani suke asircewa. Saboda modelar ta samar da rubutu mai yuwuwa jiya karbare gida, hallucination shine yanayin asali, ba occasional bug ba — don haka engineers dole ne su yi tsara kewaye. Sanin me yasa ya faru (next-token prediction tare da babu jiyya gaskiya) ya gaya maka yadda za a rage: grounding, citations, rage temperature, izini don rashin aiki, da waje tabbatar. Damuwa da LLM sakonnin azaman draft wanda a duba — ba oracle ba — shine bambanci tsakanin tamkar aiki da confident liability.
Ɗakin karatu na tambayoyin hira na IT tare da amsoshi cikakke — daga Junior zuwa Senior.
Ba da Gudummawa