paint-brush
Gukemura ikibazo cya AI Hallucination hamwe no Kwigenzura-Porogaramu Kamerena@cosmological
Amateka mashya

Gukemura ikibazo cya AI Hallucination hamwe no Kwigenzura-Porogaramu Kamere

Birebire cyane; Gusoma

Porogaramu Kamere yongerera ubwizerwe bwa LLMs mugenzura buri ntambwe yuburyo bwo gutekereza. Bitandukanye nubundi buryo, ikoresha muburyo bwo kwiga, itanga ibisobanuro bikaze kumakosa, kandi irahuza nibikorwa bidafatika. Ibi bitezimbere ubushobozi bwa AI bwo kwisuzuma no kunonosora imitekerereze idafite ibisubizo byo hanze cyangwa kurangiza.
featured image - Gukemura ikibazo cya AI Hallucination hamwe no Kwigenzura-Porogaramu Kamere
Cosmological thinking: time, space and universal causation  HackerNoon profile picture
0-item

Abanditsi:

(1) Zhan Ling, UC San Diego n'umusanzu ungana;

(2) Yunhao Fang, UC San Diego n'umusanzu ungana;

(3) Xuanlin Li, UC San Diego;

(4) Zhiao Huang, UC San Diego;

(5) Mingu Lee, Ubushakashatsi bwa Qualcomm AI hamwe nubushakashatsi bwa Qualcomm

(6) Roland Memisevic, Ubushakashatsi bwa Qualcomm;

(7) Hao Su, UC San Diego.

Imbonerahamwe

Ibisobanuro n'Intangiriro

Akazi kajyanye

Impamvu no Gutegura Ibibazo

Gukuraho Kugenzura Urunigi-rw'ibitekerezo Gutekereza

Ubushakashatsi

Imipaka

Umwanzuro, Gushimira hamwe


Kugenzura Gukuraho hamwe na Moderi ya Vicuna

B Ibiganiro Byinshi Kubijyanye no Gutezimbere Kugenzura Gukuraho Ukuri Kubinyuranye no kunoza igisubizo cyanyuma Gukosora

C Ibisobanuro birambuye kubikuramo

D Ibisobanuro

E Ibindi Byakuweho Kugenzura Ingero

2 Akazi kajyanye

Gutekereza hamwe nururimi runini. Icyitegererezo kinini cyururimi (LLMs) [3, 8, 57, 47, 38, 18, 9, 37] bagaragaje ubushobozi budasanzwe mugukemura ibibazo bigoye byo gutekereza. Aho kureka LLMs igatanga ibisubizo byanyuma nkibisohoka, imirimo ibanza yerekanye ko mugushishikariza gutekereza ku ntambwe ku yindi binyuze mu buryo bukwiye, nk'Urunigi-rw'ibitekerezo (CoT) rusaba [50] n'abandi benshi [21, 59, 58, 44, 48, 60, 25, 54], LLMs zigaragaza imikorere myiza cyane mubikorwa bitandukanye byo gutekereza. Kugirango turusheho kunoza uburyo bwo gutekereza ku ntambwe ku yindi, ubushakashatsi bumwe na bumwe buherutse gukora ubushakashatsi bwifashishije abashoramari bo hanze nk'abasobanuzi ba porogaramu [39, 5, 27], guhugura no guhamagarira ibitekerezo byo hanze [11], cyangwa gukora ubushakashatsi bweruye kugira ngo habeho intambwe zo gukuraho [2, 46]. Mugereranije niyi mirimo, ntabwo twishingikiriza kumasomo yo hanze na algorithms, kandi dukoresha muburyo butaziguye ubushobozi bwo kwiga bwa LLMs kugirango tubyare ibitekerezo byuzuye kandi bikomeye.


Ururimi runini rwicyitegererezo. Gukoresha icyitegererezo cyururimi kugirango usuzume ibisekuruza byintangarugero byabaye igitekerezo kirekire [22, 36, 40, 4]. Nkuko LLMs igaragaza ubushobozi butangaje mubikorwa bitandukanye, biba igitekerezo gisanzwe cyo gukoresha LLM nkibikoresho byo gusuzuma no kugenzura. Kurugero, [10, 11, 33] finetune LLMs kugirango igenzure ibisubizo nintambwe zo hagati. LLMs ihujwe na RLHF [32, 31, 48] nayo yakoreshejwe kugirango igereranye ibisekuru bitandukanye by'icyitegererezo. Mubyongeyeho, imirimo iherutse nka [43, 52, 28, 6] gukoresha uburyo bwihuse bwo kwemerera LLMs kwisuzuma, kwikosora, no kwikemurira ibibazo bidakenewe kurangiza. Ariko, iyi mirimo ntabwo yibanda ku gukomera no kwizerwa mubikorwa byo kugabanya ibitekerezo kuri buri ntambwe yo gutekereza. Muriyi mirimo, turasaba imvugo karemano ishingiye kumvugo yo kugabanya ibitekerezo yemerera LLMs kwisuzumisha buri ntambwe yo hagati yinzira yo gutekerezaho, bityo tunoze gukomera no kwizerwa mubitekerezo.


Imbonerahamwe 1: Ikibazo cyurugero rwa GSM8K hamwe ninzira yakozwe ya CoT yo gutekereza hamwe na GPT3.5 (turbo), aho ibisohoka bitanga urunigi rwo gutekereza nabi hamwe nigisubizo cyukuri.


Byongeye kandi, mugihe imirimo imwe n'imwe iheruka [12, 53, 15, 34] yatanze uburyo bwo kugenzura intambwe kugiti cye muburyo bwo gutekereza, inzira yacu itandukanya niyi mirimo muburyo bukurikira: (1) Uburyo bwacu bukoresha uburyo bwo kwiga kugirango tubigereho kugenzura ibitekerezo, bitabaye ngombwa ko imvugo yerekana finetuning. . . Kurugero, uburyo bwacu burahuza numurimo wanyuma wamabaruwa, aho LLM isabwa gusohora guhuza inyuguti zanyuma zamagambo yose mukurikirana nkigisubizo cyanyuma. (4) Gahunda yacu ya Kamere ituma hakoreshwa ubumenyi busanzwe butagaragara neza mubibanza. Kurugero, tekereza kuri iki kibazo: “Marin arya pome 4 kumunsi. Arya pome zingahe mu Gushyingo? ” Nubwo "Ugushyingo ifite iminsi 30" itashyizwe ku rutonde mu buryo bweruye, Porogaramu Kamere yemerera gukoresha ubumenyi nk'ubwo mu ntambwe yo gutekereza. Igenzura ryacu murwego rwo kugenzura naryo rirashoboye gukemura ibyo bibanza bitagaragara (urugero, niba LLM isohoka "Ugushyingo ifite iminsi 29" murwego rwo gutekereza, bizashyirwaho nkibitemewe).


Uru rupapuro ruraboneka kuri arxiv munsi ya CC BY 4.0 DEED.