HejuruUbuyobozi bukurikira
Nigute Ibibazo Biyobora Bias AI Ibisubizo
Ururimi AI
Ururimi AI
Logit bias ni ipfundo ryerekana urugero rwururimi rugana cyangwa kure yikimenyetso cyihariye wongeyeho umubare uhamye kumanota yabo mbere yuko icyitegererezo gitora ijambo rikurikira.
Umva iki gitaboAI-byavuzwe · ~2 min
It is a lightweight way to ban words, force choices, or shape style without retraining anything.
Mbere yuko icyitegererezo gihitamo ikimenyetso gikurikiraho, gitanga logit (amanota adasanzwe) kuri buri kimenyetso mumagambo yacyo. Logit bias igufasha kongeramo agaciro gahoraho kumurongo wibimenyetso byatoranijwe numubare wabo wibimenyetso. Kubogama kwinshi kubogamye gutuma ikimenyetso gishobora kuba icyitegererezo; kubogama kwinshi (akenshi -100 muri APIs) birabuza neza. Kuberako ihinduka riba mbere ya softmax ihindura amanota mubishoboka, ndetse kubogama kworoheje guhindura guhindura kugabana. Icy'ingenzi, kubogama ni urufunguzo rwo kwerekana indangamuntu, ntabwo ari amagambo yose - bityo ijambo-ibimenyetso byinshi rishobora gukenera buri gice cyaryo kibogamye kugirango gihagarike cyangwa kibiteze imbere. Nuburyo bwihuse, bwo kubaga budasaba kuringaniza neza kandi bukurikizwa kubisabwa.
Gushiraho -100 kubogama kubimenyetso byo gutukana kugirango wirinde chatbot guhora itanga amagambo amwe.
Guhatira yego / oya mubyiciro utanga kubogama gukomeye kubimenyetso bya 'Yego' na 'Oya' no guhagarika ibindi byose.
Guca intege interuro ikoreshwa cyane cyangwa ijambo ryuzuza ukoresheje kubogama kuringaniye kubimenyetso byayo.
Kuzamura imvugo yihariye (nkizina ryibicuruzwa) rero incamake irabivuga neza.
Kubogama kubogamye bikomeje kuba intandaro yo kuyobora byihuse, ariko ubundi buryo bukize buragenda bwiyongera: decoding yubatswe / yubatswe kuburizamo ingwate zikomeye, hamwe na enterineti ikora cyangwa ihagarariwe na injeniyeri yerekana icyerekezo cyimbere aho gutanga amanota gusa. Tegereza APIs kugumya kubogama nkibisanzwe byoroshye guhunga mugihe utanga urwego rwohejuru rwo kugenzura - interuro zabujijwe, amabwiriza yuburyo, uburyo bwo kuyungurura umutekano - ikora tokenisation mu buryo bworoshye kugirango abayitezimbere batagomba gutekereza kubijyanye nindangamuntu mbisi.
Ibaruwa y'amakuru y'ubuntu
Inkuru eshatu zizewe za AI buri gitondo cy'iminsi y'akazi, zanditse mu rurimi rworoshye. Ubuntu iteka ryose, nta kwamamaza.
Imeri imwe buri munsi y'akazi. Kwiyimura mu buryo bumwe. Ntitwigera tugurisha cyangwa ngo dusangire aderesi yawe.
Isuzume
Ibisubizo ako kanya kuri buri kibazo, n'icyemezo gishobora gusangirwa gifite ID yagenzurwa umaze gutsinda isomo.
Fasha uburezi bwa AI bw'ubuntu. AI Understanding ni umuryango utegamiye kuri leta wa 501(c)(3) — nta kwamamaza, nta kwishyura, iteka ryose. Tanga impano
Logit bias ni ipfundo ryerekana urugero rwururimi rugana cyangwa kure yikimenyetso cyihariye wongeyeho umubare uhamye kumanota yabo mbere yuko icyitegererezo gitora ijambo rikurikira. Nuburyo bworoshye bwo guhagarika amagambo, guhitamo imbaraga, cyangwa imiterere yimiterere nta kongera imyitozo.
Kubogama kubogamye byongeramo guhora kumurongo wibimenyetso byatoranijwe, uhindura uburyo bishoboka ko byatorwa, udahinduye icyitegererezo ubwacyo.
Kinini kubogama nka -100 itwara ikimenyetso cya post-softmax ishobora kuba kuri zeru, kubwibyo ntabwo yigeze ikorwa.
The API reference documents a range of -100 to 100 for each token ID, where negative values decrease likelihood and positive values increase it.
Indangagaciro zibogamye zikoreshwa kubiranga ibimenyetso byihariye biva mumagambo ya tokenizer, niyo mpamvu ugomba kumenya indangamuntu kubice byijambo.
Softmax yerekana logits, kubwibyo rero no kubogama kworoheje kubogamye kugwiza ikimenyetso kidasanzwe uburemere budasanzwe, bikazamura cyane itandukaniro ryacyo.
Komeza wige
Abandi bayobozi batoranijwe kuriyi ngingo
HejuruUbuyobozi bukurikira
Nigute Ibibazo Biyobora Bias AI Ibisubizo
Ururimi AI