Menene daidaitawa mai kyau kuma me yasa tsokanar ku ke aiki mafi kyau da shi?

Iyaka tsakanin Abin da kuke cim ma tare da saƙo mai kyau da abin da kuka samu ta hanyar daidaita ƙirar ƙira Ya fi dabara fiye da alama, amma fahimtar shi yana haifar da bambanci tsakanin martani na tsaka-tsaki da tsarin amfani na gaske. A cikin wannan jagorar, zan nuna muku, tare da misalai da kwatancen, yadda za ku zaɓa da haɗa kowace dabara don cimma ingantaccen sakamako a cikin ayyukan zahiri na duniya.

Makasudin ba shine ka tsaya a cikin ka'idar ba, amma don aiwatar da shi a kullun: lokacin da aikin injiniya da sauri ya ishe ku, Yaushe ya cancanci saka hannun jari a cikin daidaitawa mai kyau?, Ta yaya duk wannan ya dace da raƙuman ruwa na RAG, da kuma waɗanne ayyuka mafi kyau na rage farashi, haɓaka haɓakawa, da kuma guje wa shiga cikin matattu.

Menene aikin injiniya mai sauri, saurin daidaitawa, da kuma daidaitawa mai kyau?

Kafin mu ci gaba, bari mu fayyace wasu dabaru:

• Injiniyan gaggawa shine fasaha na zayyana bayyanannun umarni tare da ingantattun mahallin da tsammanin. don jagorantar samfurin da aka riga aka horar. A cikin a chatbot, alal misali, yana bayyana rawar, sautin, tsarin fitarwa, da misalai don rage rashin fahimta da inganta daidaito ba tare da taɓa ma'aunin samfurin ba.
• Kyakkyawan daidaitawa yana canza sigogi na ciki na ƙirar da aka riga aka horar tare da ƙarin bayanai daga yankin. don daidaita ayyukanku akan takamaiman ayyuka. Yana da manufa lokacin da kuke buƙatar ƙwararrun kalmomi, shawarwari masu rikitarwa, ko matsakaicin daidaito a cikin yankuna masu mahimmanci (kiwon lafiya, doka, kuɗi).
• Gyaran gaggawa yana ƙara ƙwararrun ƙwararru (mai laushi mai laushi) waɗanda ƙirar ke fassara tare da rubutun shigarwaBa ya sake horar da duka samfurin: yana daskare nauyinsa kuma yana inganta kawai waɗanda aka saka "waƙoƙi." Yana da ingantaccen tsaka-tsaki lokacin da kake son daidaita hali ba tare da farashin cikakken daidaitawa ba.

A cikin ƙirar UX/UI, aikin injiniya mai sauri yana inganta ingantaccen hulɗar ɗan adam-kwamfuta (abin da nake tsammani da kuma yadda nake nema), yayin da daidaitawa yana ƙara dacewa da daidaiton fitarwa. Hade, ba da izini don ƙarin amfani, sauri, kuma amintattun musaya.

Injiniya mai sauri a cikin zurfin: dabarun da ke motsa allura

Injiniyan gaggawa ba batun gwajin makaho bane. Akwai hanyoyin tsari wanda ke inganta inganci ba tare da taɓa samfurin ko bayanan tushe ba:

• 'Yan-shot vs sifili-shot. a 'yan-harbi Kuna ƙara wasu misalai da aka zaɓa da kyau don samfurin ya ɗauki ainihin tsari; in sifili-harba Kuna dogara da takamaiman umarni da haraji ba tare da misalai ba.
• Zanga-zangar a cikin mahallin. Nuna tsarin da ake sa ran (shigarwa → fitarwa) tare da ƙananan nau'i-nau'i. Wannan yana rage kurakuran tsarawa kuma yana daidaita tsammanin, musamman idan kuna buƙatar takamaiman filayen, lakabi, ko salo a cikin martani.
• Samfura da masu canjiƘayyade tsokaci tare da masu riƙe wuri don canza bayanai. Ƙwararru mai ƙarfi shine maɓalli lokacin da tsarin shigarwa ya bambanta, misali, ta hanyar tsaftace bayanai ko gogewa inda kowane rikodin ya zo a cikin wani tsari daban.
• Masu maganaSu ne "masu fassara" tsakanin sararin rubutu na samfurin da nau'ikan kasuwancin ku (misali, taswirar "mai farin ciki" → "tabbatacce"). Zaɓin masu magana mai kyau yana inganta daidaiton lakabi da daidaito, musamman a cikin nazarin ra'ayi da rabe-raben jigogi.
• Wuraren gaggawa (sarkar gaggawa). Rarraba hadadden aiki zuwa matakai: taƙaita → Cire ma'auni → nazarin jin daɗi. Sarkar matakai tare yana sa tsarin ya zama mai lalacewa da ƙarfi, kuma sau da yawa yana inganta inganci idan aka kwatanta da "neman komai a lokaci ɗaya."
• Kyawawan ayyukan tsarawa: alamar matsayi ("Kai mai nazari ne..."), ya bayyana salon ("amsa a cikin tebur / JSON"), ya kafa ma'auni na kimantawa ("yana azabtar da hallucinations, ya ambaci tushe lokacin da suke wanzu") kuma ya bayyana abin da za a yi idan akwai rashin tabbas (misali, "idan bayanan ya ɓace, nuna 'ba a sani ba'").

Abubuwan gyarawa da sauri

Bugu da ƙari ga faɗakarwa na yanayi, saurin kunnawa yana haɗa da tsokaci mai laushi (nau'i-nau'i masu horarwa) waɗanda ke gaban shigarwar. Yayin horo, gradient yana daidaita waɗancan hanyoyin don kawo abin da ake fitarwa kusa da manufa. ba tare da shafar samfurin sauran nauyin nauyi ba. Yana da amfani lokacin da kuke son ɗaukar hoto da ƙananan farashi.

Kuna loda LLM (misali, GPT-2 ko makamancin haka), shirya misalan ku da kuna shirya abubuwan motsa jiki masu taushi don kowane shigarwaKuna horar da waɗancan abubuwan haɗawa ne kawai, don haka ƙirar tana “gani” ingantaccen gabatarwar da ke jagorantar halayen sa a cikin aikinku.

Aikace-aikace na aiki: A cikin bot ɗin sabis na abokin ciniki, zaku iya haɗawa da tsarin tambaya na yau da kullun da ingantaccen sautin amsawa a cikin tsokaci mai laushi. Wannan yana hanzarta daidaitawa ba tare da kiyaye rassan samfura daban-daban ba. ko cinye ƙarin GPU.

Tunatarwa mai zurfi mai zurfi: yaushe, ta yaya, kuma da wane taka tsantsan

Kyakkyawan daidaitawa na sake horar da ma'aunin LLM tare da saitin bayanai masu niyya. don kware shi. Wannan ita ce hanya mafi kyau idan aikin ya kauce daga abin da samfurin ya gani a lokacin horo na farko ko kuma yana buƙatar ƙayyadaddun kalmomi da yanke shawara.

Ba za ku fara daga faifai ba: samfuran da aka daidaita taɗi kamar gpt-3.5-turbo An riga an kunna su don bin umarni. Sauraron ku mai kyau yana "amsa" ga wannan hali, wanda zai iya zama da hankali da rashin tabbas, don haka yana da kyau a yi gwaji tare da ƙirar tsarin da aka sawa da shigarwa.

Wasu dandamali suna ba ku damar ɗaure sauti mai kyau akan na yanzu. Wannan yana ƙarfafa sigina masu amfani a ƙananan farashi. don ja da baya daga karce, kuma yana sauƙaƙe ƙwaƙƙwaran ingantattun gyare-gyare.

Ingantattun dabaru irin su LoRA saka ƙananan matrices don daidaita ƙirar tare da ƴan sabbin sigogi. Riba: ƙananan amfani, agile turawa da sake juyawa (zaku iya "cire" daidaitawa ba tare da taɓa tushe ba).

Kwatanta: saurin kunnawa vs kyakkyawan kunnawa

• Tsarin aikiKyakkyawan daidaitawa yana sabunta ma'aunin ƙira tare da alamar bayanan manufa; Gyaran gaggawa yana daskare samfurin kuma yana daidaita abubuwan da za a iya horarwa kawai waɗanda aka haɗa su zuwa shigarwar; injiniya mai sauri yana inganta rubutun koyarwa da misalai marasa horo.
• Saitin sigogiA cikin daidaitawa mai kyau, kuna canza hanyar sadarwar; a cikin saurin kunnawa, kawai kuna taɓa "tallafi mai laushi." A cikin aikin injiniya na gaggawa, babu daidaitawa, kawai ƙira.
• Tsarin shigarwaKyakkyawan daidaitawa yawanci yana mutunta tsarin asali; saurin daidaitawa yana sake fasalin shigarwa tare da haɗawa da samfuri; aikin injiniya mai sauri yana ba da damar ingantaccen harshe na halitta (matsayi, ƙuntatawa, misalai).
• ResourcesKyakkyawan daidaitawa ya fi tsada (ƙididdigewa, bayanai, da lokaci); saurin daidaitawa ya fi dacewa; Injiniya mai sauri shine mafi arha kuma mafi sauri don maimaitawa idan harka ta yarda.
• Manufar da kasadaKyakkyawan daidaitawa yana inganta kai tsaye zuwa aikin, yana kawar da haɗarin wuce gona da iri; saurin daidaitawa ya yi daidai da abin da aka riga aka koya a cikin LLM; aikin injiniya mai sauri yana rage hasashe da kurakurai masu tsarawa tare da mafi kyawun ayyuka ba tare da taɓa samfurin ba.

Bayanai da kayan aiki: man fetur na aikin

• Ingancin bayanai na farko: waraka, deduplication, daidaitawa, gefen yanayin ɗaukar hoto da metadata masu yawa Su ne kashi 80% na sakamakon, ko kuna yin gyaran-gyare ko gyara gaggawa.
• Mai sarrafa bututun mai: dandamalin injiniyan bayanai don haɓaka AI (misali, mafita waɗanda ke ƙirƙirar samfuran bayanan da za a sake amfani da su) taimaka haɗawa, canzawa, isarwa da saka idanu bayanan bayanan domin horo da kimantawa. Ra'ayoyi kamar "Nexsets" suna kwatanta yadda ake shirya bayanan da aka shirya don amfani da ƙira.
• Madauki na martani: Tattara siginonin amfani na zahiri (nasara, kurakurai, tambayoyin da ake yawan yi) kuma a mayar da su cikin abubuwan faɗakarwa, faɗakarwa mai laushi, ko saitin bayanai. Ita ce hanya mafi sauri don samun daidaito.
• sake haifuwa: Sigar faɗakarwa, faɗakarwa mai laushi, bayanai, da ma'aunin ma'auni masu dacewa. Ba tare da ganowa ba, ba zai yuwu a san abin da ya canza aikin ba ko kuma komawa cikin yanayi mai kyau idan haɓakawar ta gaza.
• IzationaddamarwaLokacin fadada ɗawainiya ko harsuna, tabbatar da masu faɗar magana, misalan ku, da lakabinku ba a keɓance su da wani takamaiman yanki ba. Idan kuna canza a tsaye, ƙila kuna buƙatar yin wasu gyarar haske ko amfani da sabbin tausasawa.
• Me zai faru idan na canza tambayar bayan gyarawa? Gabaɗaya, a: ƙirar yakamata ta bambanta salo da ɗabi'a daga abin da aka koya, ba kawai maimaita alamu ba. Wannan shine ainihin ma'anar ingin inference.
• Rufe madauki tare da ma'auniBayan daidaito, yana auna daidaitaccen tsari, ɗaukar hoto, ambaton tushe a cikin RAG, da gamsuwar mai amfani. Abin da ba a auna ba ya inganta.

Zaɓi tsakanin faɗakarwa, saurin daidaitawa da daidaitawa ba batun akida bane amma na mahallin.: farashi, ma'auni na lokaci, haɗarin kuskure, samuwar bayanai, da buƙatar ƙwarewa. Idan kun ƙusa waɗannan abubuwan, fasaha za ta yi aiki a cikin yardar ku, ba ta wata hanya ba.

Daniel Terrasa

Edita ya ƙware a fannin fasaha da al'amuran intanet tare da gogewa fiye da shekaru goma a cikin kafofin watsa labaru na dijital daban-daban. Na yi aiki a matsayin edita da mahaliccin abun ciki don kasuwancin e-commerce, sadarwa, tallan kan layi da kamfanonin talla. Na kuma yi rubutu a shafukan yanar gizo na tattalin arziki, kudi da sauran fannoni. Aikina kuma shine sha'awata. Yanzu, ta hanyar labarai na a ciki Tecnobits, Ina ƙoƙarin bincika duk labarai da sababbin damar da duniyar fasahar ke ba mu kowace rana don inganta rayuwarmu.