Okpomọkụ na Sampling
Okpomọkụ na nlele bụ ọpịpị na-achịkwa ka 'random' ma ọ bụ 'dịkwakwa' nke okwu ụdị asụsụ si dị.
Nchịkọta
They decide whether you get the same predictable answer every time or fresh, varied phrasing.
Ime miri emi
Na nzọụkwụ ọ bụla, ụdị asụsụ anaghị ewepụta okwu ozugbo - ọ na-ewepụta akara ('logit') maka akara ọ bụla n'ime okwu ya, nke softmax na-atụgharị na nkesa nke puru omume. Nlere anya bụ ka esi ahọrọ akara na-esote site na nkesa ahụ. Okpomọkụ na-emegharị nkesa tupu ịhọrọ: obere okpomọkụ na-eme ka nhọrọ ndị dị elu na-achịkwa, ya mere mmepụta na-elekwasị anya ma na-emegharịghachi; okpomọkụ dị elu na-eme ka ọ daa, na-eme ka akara ngosi ndị na-atụghị anya ya banye maka ụdị dịgasị iche iche (na ọtụtụ njehie). Ihe nzacha abụọ ewu ewu na-ebu ụzọ mechie ọdọ mmiri ahụ. Top-k na-edobe naanị akara puru omume k kacha elu. Top-p, ma ọ bụ nucleus sampling, na-edobe nke kacha nta nke tokens nke puru omume gbakwunyere na p (kwuru 0.9), otú ọdọ mmiri na-eto eto mgbe ihe nlereanya na-ejighị n'aka na-ebelata mgbe ọ nwere obi ike. Ngwakọta ntọala ndị a na-agbanwe ntụkwasị obi megide okike.
Nghọta nka nka
Okpomọkụ na-arụ ọrụ site n'ịkewa logt ọ bụla site na T tupu softmax: ihe gbasara nke puru omume bụ nke dabara na exp(logit / T). T n'okpuru 1 na-amụba oghere ka akara elu na-achịkwa; T n'elu 1 na-ebelata oghere ndị ahụ ma gbasaa nkesa. Na T nso 0 ihe nlereanya ahụ na-aghọ onye anyaukwu nke ọma, na-ewere akara ngosi nke nwere ike ịbụ mgbe niile. Top-k na-ekpuchi onye ndoro-ndoro na ọnụ ọgụgụ a kapịrị ọnụ, ebe top-p na-esetịpụ nkwụsị nke puru omume, ya mere ọnụ ọgụgụ ndị na-akwado ya na-adaba n'otu obi siri ike na ihe nlereanya ahụ dị na nzọụkwụ ahụ.
Mmetụta atụmatụ
Ọsọ na ọnụ ọgụgụ
Usoro ọrụ asụsụ nwere ike ịga ngwa ngwa n'achụghị nkwụsi ike.
Nweta na iru
Ọ na-agbasawanye ohere n'ofe asụsụ na ụdị nzikọrịta ozi.
Mkpebi doro anya
Ndị otu nwere ike itinyekwu oge na ikpe ebe akpaaka na-ejikwa nkwughachi.
Ọdịnihu nke okpomọkụ na nlele
Njikwa ndị a kwụsiri ike ma ghọta nke ọma, yabụ omume a dị na ndabara mara mma yana ụdị ọhụrụ. Na-atụ anya atụmatụ mmeghari ndị ọzọ dị ka min-p (nke na-eme ka mbelata ahụ ruo n'ike nke token dị elu) yana ọnọdụ okpomọkụ na-agbanwe n'etiti ọgbọ. Ngwá ọrụ ga-arịbawanye ntọala akpaaka maka ọrụ ọ bụla - dị ala maka koodu na mmịpụta, dị elu maka ịtụgharị uche - yabụ ndị ọrụ agaghị eji aka na-akụ. Isi echiche na-atachi obi: nlele bụ ihe dị mfe, dị ike n'etiti njiri mara nkenke na ụdị okike dị iche iche.
Mmejuputa n'ezie n'ụwa
Tọọ okpomọkụ nso 0 maka ọgbọ koodu ma ọ bụ mmịpụta data, ebe ịchọrọ otu azịza ziri ezi oge ọ bụla
Na-ebuli okpomọkụ ruo ihe dị ka 0.8-1.0 maka ịtụgharị uche aha, slogans, ma ọ bụ echiche akụkọ iji nweta nhọrọ dị iche iche
Iji top-p gburugburu 0.9 ya mere ihe atụ ahụ si na naanị okwu ndị nwere ezi uche ma na-ezere akara ngosi dị egwu.
Itinye top-k na ndị na-achọ okpu ma gbochie okwu ndị na-adịghị ahụkebe na-adịghị ahụkebe ịpụta na nzaghachi na-eche ndị ahịa ihu.
Ihe ize ndụ & okporo ụzọ nche
Eziokwu ndị e chepụtara echepụta nwere ike jiri nwayọ tinye akụkọ, nkwado nkwado, ma ọ bụ nsonaazụ nyocha.
Mmetụta ngwa ngwa nwere ike ịmepụta nsonaazụ na-ekwekọghị ekwekọ n'ofe arịrịọ ndị yiri ya.
Enwere ike ikpughe data ederede nwere mmetụta ma ọ bụrụ na njikwa ohere adịghị ike.
Map mmejuputa
Kọwaa usoro mmepụta, ụda, na ụkpụrụ ịdịmma tupu ibugharị.
Weghachite nzaghachi site na isi mmalite ntụkwasị obi mgbe ọ bụla izi ezi dị mkpa.
Debe ebe nleba anya mmadụ maka mpụta dị elu.
Sochie ụkpụrụ ọdịda ma na-azụghachi mkpali ma ọ bụ usoro ọrụ mgbe niile.
Nọgide na-eme nchọpụta
Free newsletter
Get the daily AI briefing
Three verified AI stories every weekday morning, written in plain English. Free forever, no ads.
One email each weekday. Unsubscribe in one click. We never sell or share your address.
Test yourself
Take the Temperature and Sampling quiz
Instant feedback on every answer, and a shareable certificate with a verifiable ID once you pass a course.
Support free AI education. AI Understanding is a 501(c)(3) nonprofit — no ads, no paywall, ever. Make a donation
Ntuziaka na-esote
Nucleus na Top-k Sampling
Ajụjụ a na-ajụkarị
What is Temperature and Sampling?
Okpomọkụ na nlele bụ ọpịpị na-achịkwa ka 'random' ma ọ bụ 'dịkwakwa' nke okwu ụdị asụsụ si dị. Ha na-ekpebi ma ị na-enweta otu azịza a ga-ebu amụma oge ọ bụla ma ọ bụ nke dị ọhụrụ, nkebiokwu dị iche iche.
Kedu ihe iweli okpomọkụ na-eme na mmepụta ihe nlereanya?
Okpomọkụ dị elu na-eme ka nkesa nwere ike gbasaa, yabụ, a na-ewerekarị akara ndị na-adịkarị obere, na-ewepụta mpụta dị iche iche (ma dị ize ndụ).
Kedu ka nlele top-p (nucleus) si dị iche na nlele top-k?
Top-p na-emegharị onye ndoro-ndoro anya site na nchikota nke puru omume, ebe top-k na-edobe ọnụ ọgụgụ akara ngosi kachasị elu mgbe niile.
Mechanically, kedu ihe okpomọkụ na-eme n'ezie na logit?
Okpomọkụ na-ekewa logt ọ bụla site na T tupu softmax; T n'okpuru 1 na-amụba nkesa, T n'elu 1 na-eme ka ọ dị nro.
Maka imepụta koodu ma ọ bụ wepụta data ahaziri ahazi, kedu ntọala kacha mma?
Ọrụ ndị chọrọ nzizi na nkwugharị na-eji oke okpomọkụ dị ala nke mere na ihe nlereanya ahụ na-ahọrọ nke kachasị, akara ngosi ziri ezi.
Kedu ihe na-eme na okpomọkụ nke 0 nke ọma?
Na nso okpomọkụ efu, nkesa dị nkọ nke na a na-ahọrọ akara ngosi kachasị elu mgbe niile - ngbanwe anyaukwu.