M'ndandanda wazopezekamo[Bisani][Show]
Vuto lachikale mu luntha lochita kupanga ndilo kufunafuna makina omwe amatha kumvetsetsa chilankhulo cha anthu.
Mwachitsanzo, pofufuza "malo odyera aku Italiya apafupi" pakusaka kwanu komwe mumakonda, njira yowerengera imayenera kusanthula liwu lililonse mufunso lanu ndikutulutsa zotsatira zoyenera. Pulogalamu yabwino yomasulira iyenera kumvetsetsa tanthauzo la liwu linalake mu Chingerezi ndipo mwanjira ina imatengera kusiyana kwa galamala pakati pa zilankhulo.
Ntchito zonsezi ndi zina zambiri zimagwera pansi pa gawo la sayansi yamakompyuta yotchedwa Kusintha kwa Chilengedwe Chachilengedwe kapena NLP. Kupita patsogolo kwa NLP kwapangitsa kuti pakhale ntchito zingapo zothandiza kuchokera kwa othandizira ngati Amazon's Alexa kupita ku zosefera za spam zomwe zimazindikira maimelo oyipa.
Kupambana kwaposachedwa kwambiri mu NLP ndi lingaliro la a chilankhulo chachikulu kapena LLM. Ma LLM monga GPT-3 akhala amphamvu kwambiri kotero kuti akuwoneka kuti akuchita bwino pafupifupi ntchito iliyonse ya NLP kapena kugwiritsa ntchito.
M'nkhaniyi, tiwona zomwe ma LLM ali, momwe mitunduyi imaphunzitsidwira, komanso zofooka zomwe ali nazo.
Kodi chilankhulo chachikulu ndi chiyani?
Pachimake chake, chilankhulo cha chilankhulo ndi algorithm yokha yomwe imadziwa kuti kutsatizana kwa mawu kuli chiganizo chovomerezeka.
Chilankhulo chosavuta kwambiri chophunzitsidwa pamabuku mazana angapo chiyenera kudziwa kuti "Anapita kunyumba" ndi chovomerezeka kuposa "Kunyumba anapita".
Ngati tisintha deta yaying'ono ndi dataset yayikulu yochotsedwa pa intaneti, timayamba kuyandikira lingaliro la a chilankhulo chachikulu.
kugwiritsa mawindo a neural, ofufuza akhoza kuphunzitsa LLMs pa chiwerengero chachikulu cha malemba. Chifukwa cha kuchuluka kwa zolemba zomwe chitsanzocho chawona, LLM imakhala yabwino kwambiri pakulosera mawu otsatirawa motsatizana.
Mtunduwu umakhala wovuta kwambiri, umatha kugwira ntchito zambiri za NLP. Ntchitozi zikuphatikiza kufotokoza mwachidule zolemba, kupanga zolemba zatsopano, komanso kuyerekezera zokambirana ngati za anthu.
Mwachitsanzo, chilankhulo chodziwika bwino cha GPT-3 chimaphunzitsidwa ndi magawo opitilira 175 biliyoni ndipo chimawerengedwa kuti ndicho chiyankhulo chapamwamba kwambiri mpaka pano.
Itha kupanga code yogwira ntchito, kulemba zolemba zonse, ndipo imatha kuwombera mafunso okhudza mutu uliwonse.
Kodi ma LLM Amaphunzitsidwa Motani?
Takhudza mwachidule mfundo yakuti ma LLM ali ndi mphamvu zambiri chifukwa cha kukula kwa maphunziro awo. Pali chifukwa chomwe timachitcha kuti "zachiyankhulo zazikulu" pambuyo pake.
Maphunziro asanachitike ndi Transformer Architecture
Panthawi yophunzitsira, ma LLM amadziwitsidwa pazomwe zilipo kale kuti aphunzire kalembedwe ndi malamulo achilankhulo.
M'zaka zingapo zapitazi, ma LLM adaphunzitsidwa kale pama dataset omwe amakhudza gawo lalikulu la intaneti. Mwachitsanzo, chilankhulo cha GPT-3 chinaphunzitsidwa pa data kuchokera ku Kukwawa Kwamba dataset, kuchuluka kwa zolemba, masamba, ndi mabuku osungidwa pakompyuta omwe adachotsedwa m'madomeni opitilira 50 miliyoni.
Gulu lalikulu la deta limasinthidwa kukhala chitsanzo chodziwika kuti a transformer. Transformers ndi mtundu wa neural network yakuya zomwe zimagwira ntchito bwino pama data otsatizana.
Transformers amagwiritsa ntchito kamangidwe ka encoder-decoder za kusamalira zolowetsa ndi zotuluka. Kwenikweni, thiransifoma imakhala ndi ma neural network awiri: encoder ndi decoder. Encoder imatha kutulutsa tanthauzo la mawu olowera ndikusunga ngati vekitala. Kenako decoder imalandira vector ndikupanga kumasulira kwake kwa mawuwo.
Komabe, lingaliro lofunikira lomwe linalola kuti mapangidwe a transformer agwire ntchito bwino ndikuwonjezera a njira yodzisamalira. Lingaliro la kudzisamalira linalola chitsanzo kumvetsera mawu ofunika kwambiri mu chiganizo choperekedwa. Kachipangizoka kamayang'ananso kulemera kwa mawu omwe amasiyana kwambiri motsatizana.
Phindu lina la kudzisamalira ndilokuti ndondomekoyi ikhoza kufananizidwa. M'malo mokonza deta yotsatizana, zitsanzo za transformer zimatha kukonza zolowetsa zonse nthawi imodzi. Izi zimathandiza ma transfoma kuti aphunzitse zambiri za data mwachangu poyerekeza ndi njira zina.
Sintha pang'onopang'ono
Pambuyo pa gawo lokonzekera, mutha kusankha kuyambitsa mawu atsopano a LLM yoyambira. Timatcha ndondomekoyi sintha pang'onopang'ono ndipo nthawi zambiri amagwiritsidwa ntchito kupititsa patsogolo kutulutsa kwa LLM pa ntchito inayake.
Mwachitsanzo, mungafune kugwiritsa ntchito LLM kupanga zomwe zili pa akaunti yanu ya Twitter. Titha kupatsa chitsanzochi ndi zitsanzo zingapo za ma tweets anu am'mbuyomu kuti tikupatseni lingaliro la zomwe mukufuna.
Pali mitundu ingapo yosinthira bwino.
Kuphunzira pang'ono amatanthauza njira yoperekera chitsanzo chiwerengero chochepa cha zitsanzo ndikuyembekeza kuti chitsanzo cha chinenero chidzazindikira momwe angapangire zotsatira zofanana. Kuphunzira kamodzi ndi njira yofananira kupatula chitsanzo chimodzi chokha chaperekedwa.
Zoperewera za Zinenero Zazinenelo Zazikulu
Ma LLM monga GPT-3 amatha kuchita zambiri zogwiritsa ntchito ngakhale osakonza bwino. Komabe, zitsanzozi zimabwerabe ndi malire awo.
Kupanda Kumvetsetsa kwa Semantic Padziko Lonse
Pamwamba, ma LLM amawoneka kuti akuwonetsa luntha. Komabe, zitsanzozi sizigwira ntchito mofananamo ubongo waumunthu amachita. Ma LLM amangodalira kuwerengera kuti apange zotuluka. Alibe mphamvu yolingalira paokha malingaliro ndi malingaliro.
Chifukwa cha izi, LLM imatha kutulutsa mayankho opanda pake chifukwa mawuwo amawoneka ngati "olondola" kapena "mwachiwerengero" akayikidwa mu dongosolo lomwelo.
Hallucinations
Mitundu ngati GPT-3 imakhalanso ndi mayankho olakwika. Ma LLM amatha kuvutika ndi chodabwitsa chomwe chimadziwika kuti kuyerekezera kumene zitsanzo zimatulutsa yankho lolakwika popanda kuzindikira kuti yankho lilibe maziko enieni.
Mwachitsanzo, wogwiritsa ntchito angafunse chitsanzo kuti afotokoze maganizo a Steve Jobs pa iPhone yatsopano. Chitsanzochi chikhoza kupanga mawu kuchokera ku mpweya wochepa kwambiri potengera zomwe amaphunzitsidwa.
Zokondera ndi Chidziwitso Chochepa
Mofanana ndi ma algorithms ena ambiri, mitundu yayikulu ya zilankhulo imakonda kutengera kukondera komwe kulipo pamaphunziro. Pamene tikuyamba kudalira kwambiri ma LLM kuti tipeze zambiri, okonza zitsanzozi ayenera kupeza njira zochepetsera zotsatira zovulaza za mayankho okondera.
Momwemonso, mawonekedwe akhungu a deta yophunzitsira yachitsanzo adzalepheretsanso chitsanzocho. Panopa, zitsanzo zazikulu za zinenero zimatenga miyezi kuti ziphunzitse. Zitsanzozi zimadaliranso ma dataset omwe ali ndi malire. Ichi ndichifukwa chake ChatGPT imangokhala ndi chidziwitso chochepa cha zochitika zomwe zidachitika mchaka cha 2021.
Kutsiliza
Zilankhulo zazikulu zimatha kusintha momwe timalumikizirana ndiukadaulo komanso dziko lathu lonse.
Deta yochuluka yomwe ilipo pa intaneti yapatsa ofufuza njira yowonetsera zovuta za chinenero. Komabe, panjira, zitsanzo za zilankhulo izi zikuwoneka kuti zatengera kumvetsetsa kwa dziko lapansi monga momwe zilili.
Pamene anthu ayamba kukhulupirira zitsanzo za zilankhulo izi kuti zipereke zotsatira zolondola, ochita kafukufuku ndi okonza mapulogalamu akupeza kale njira zowonjezera zowonjezera kuti teknoloji ikhalebe yolondola.
Kodi tsogolo la ma LLM ndi lotani?
Siyani Mumakonda