Dambudziko rekare muhungwaru hwekugadzira nderekutsvaga muchina unogona kunzwisisa mutauro wevanhu.
Semuyenzaniso, kana uchitsvaga "maresitorendi eItaly ari padyo" painjini yako yekutsvaga yaunofarira, algorithm inofanirwa kuongorora izwi rega rega mumubvunzo wako uye kuburitsa mhinduro dzakakodzera. Shanduro yakanaka yeapp ichafanirwa kunzwisisa mamiriro erimwe izwi muChirungu uye neimwe nzira inokonzeresa mutsauko wegirama pakati pemitauro.
Ese mabasa aya uye nezvimwe zvakawanda zvinowira pasi peiyo subfield yesainzi yekombuta inozivikanwa se Natural Language Processing kana NLP. Kufambira mberi muNLP kwakatungamira kune akawanda akawanda anoshanda maapplication kubva kune chaiwo vabatsiri seAmazon Alexa kune spam mafirita anoona ine hutsinye email.
Kubudirira kwakanyanya muNLP ipfungwa ye muenzaniso mukuru wemutauro kana LLM. MaLLM akadai seGPT-3 ave ane simba zvekuti anoita kunge ari kubudirira mune chero basa reNLP kana kese yekushandisa.
Muchinyorwa chino, isu tichatarisa kuti chii chaizvo maLLMs, kuti mamodheru aya anodzidziswa sei, uye zvipimo zvazvino zvavanazvo.
Chii chinonzi chimiro chikuru chemutauro?
Pakati payo, modhi yemutauro inongova algorithm inoziva kuti kutevedzana kwemazwi kungangoita sei mutsara unoshanda.
Mutevedzeri wemutauro wakapfava wakadzidziswa pamazana mashoma emabhuku unofanira kukwanisa kutaura kuti "Akaenda kumba" inoshanda kupfuura "Kumba akaenda".
Kana tikatsiva iyo diki dataset nehombe dataset yakarukwa kubva painternet, tinotanga kusvika kune iyo pfungwa ye muenzaniso mukuru wemutauro.
kushandisa neural networks, vatsvakurudzi vanogona kudzidzisa LLMs pahuwandu hwemashoko ezvinyorwa. Nekuda kwehuwandu hwe data data iyo modhi yakaona, iyo LLM inova yakanaka kwazvo kufanotaura izwi rinotevera mukutevedzana.
Iyo modhi inova yakaoma kwazvo, inogona kuita akawanda eNLP mabasa. Aya mabasa anosanganisira kupfupisa zvinyorwa, kugadzira zvinyorwa zvitsva, uye kunyange kutevedzera-sekutaura kwevanhu.
Semuenzaniso, iyo inonyanya kufarirwa yeGPT-3 mutauro modhi inodzidziswa neanopfuura 175 bhiriyoni paramita uye inoonekwa seyo yakanyanya kukwirisa mutauro modhi kusvika parizvino.
Inokwanisa kugadzira kodhi yekushanda, kunyora zvinyorwa zvese, uye inogona kutora pfuti pakupindura mibvunzo nezve chero musoro.
MaLLM Anodzidziswa Sei?
Tabata muchidimbu nyaya yekuti maLLM ane chikwereti chesimba rawo rakawanda kuhukuru hwedhata ravo rekudzidzisa. Pane chikonzero nei tichivadaidza kuti "makuru" emitauro yemitauro mushure mezvose.
Pre-kudzidziswa neTransformer Architecture
Munguva yepre-training nhanho, maLLM anounzwa kune iripo data data kuti vadzidze chimiro chakajairwa nemitemo yemutauro.
Mumakore mashoma apfuura, LLMs dzakafanodzidziswa pamadatasets anovhara chikamu chakakosha cheruzhinji internet. Semuenzaniso, GPT-3's mutauro modhi yakadzidziswa pane data kubva ku Yakajairika Kukamba dhatabheti, korasi yezvinyorwa zvepawebhu, mapeji ewebhu, uye mabhuku edigital akarukwa kubva munzvimbo dzinopfuura 50 miriyoni.
Iyo dataset yakakura inozopihwa mumodeli inozivikanwa se transformer. Transformers imhando ye yakadzika neural network iyo inoshanda zvakanyanya kune sequential data.
Transformers inoshandisa an encoder-decoder architecture pakubata kupinza uye kubuda. Chaizvoizvo, iyo transformer ine maviri neural network: encoder uye decoder. Iyo encoder inogona kubvisa zvinoreva zvinyorwa zvekupinza uye kuichengeta sevector. Decoder yobva yagashira vhekita yoburitsa dudziro yayo yechinyorwa.
Nekudaro, iyo yakakosha pfungwa yakabvumira iyo transformer dhizaini kuti ishande zvakanaka ndeyekuwedzerwa kwe kuzvidzora maitiro. Pfungwa yekuzvitarisa yakabvumira muenzaniso kuti uteerere kumashoko anonyanya kukosha mumutsara wakapiwa. Muchina wacho unotofunga nezve huremu huri pakati pemazwi akaparadzana nekutevedzana.
Imwe bhenefiti yekuzvitarisa ndeyekuti maitiro anogona kufananidzwa. Panzvimbo pekugadzirisa zvakatevedzana data muhurongwa, mamodhi eshanduko anogona kugadzirisa zvese zvinopinda kamwechete. Izvi zvinoita kuti vashanduri vadzidzise pahuwandu hwe data nekukurumidza zvichienzaniswa nedzimwe nzira.
Kugadzirisa zvakanaka
Mushure mechikamu che-pre-training, unogona kusarudza kusuma chinyorwa chitsva cheiyo LLM yekudzidzira. Isu tinodaidza nzira iyi kunyatsogadzirisa uye inowanzo shandiswa kusimudzira kubuda kweLLM pane rimwe basa.
Semuenzaniso, ungangoda kushandisa LLM kugadzira zvemukati zve Twitter account yako. Isu tinogona kupa iyo modhi nemienzaniso yakati wandei yematweets ako apfuura kuti tipe iyo pfungwa yezvaunoda kuburitsa.
Kune marudzi mashoma akasiyana ekugadzirisa zvakanaka.
Kudzidza zvishoma zvinoreva nzira yekupa muenzaniso nhamba diki yemienzaniso ine tarisiro yekuti chimiro chemutauro chichaziva mabudiro akafanana. Imwe-pfuti kudzidza muitiro wakafanana kunze kwemuenzaniso mumwe chete unopiwa.
Kuganhurirwa kweMienzaniso Yakakura Mutauro
MaLLM akadai seGPT-3 anokwanisa kuita nhamba huru yemakesi ekushandisa kunyangwe pasina kutsetseka. Nekudaro, aya mamodheru achiri kuuya neawo seti yekugumira.
Kushaya Kunzwisisa Semantic Kwenyika
Pamusoro, maLLM anoita seanoratidza hungwaru. Nekudaro, mamodheru aya haashande nenzira imwechete ubongo hwevanhu anoita. MaLLM anongovimba nenhamba dzemakomputa kuti abudise zvinobuda. Havana simba rekufunga mazano uye pfungwa pachavo.
Nekuda kweizvi, LLM inogona kuburitsa mhinduro dzisina musoro nekuda kwekuti mazwi anoita se "akarurama" kana "zviverengero zvingangoitika" kana akaiswa mune iyoyo hurongwa.
Hallucinations
Mienzaniso yakaita seGPT-3 inotamburawo nemhinduro dzisina kururama. MaLLM anogona kutambura nechiitiko chinozivikanwa se kuona zvinhu zvisipo uko mamodheru anoburitsa mhinduro isiriyo yechokwadi pasina kuziva kuti mhinduro haina hwaro muchokwadi.
Semuenzaniso, mushandisi anogona kubvunza modhi kuti atsanangure pfungwa dzaSteve Jobs pane yazvino iPhone. Iyo modhi inogona kuburitsa quote kubva kumhepo yakatetepa zvichibva pane yayo yekudzidziswa data.
Rusarura uye Ruzivo Rushoma
Kufanana nemamwe akawanda maalgorithms, mhando dzemitauro mikuru dzinowanzogara nhaka yekusarura iripo mune data rekudzidziswa. Apo patinotanga kuvimba zvakanyanya neLLMs kuti titore ruzivo, vanogadzira mamodheru aya vanofanirwa kutsvaga nzira dzekudzikisa zvingangokuvadza zvemhinduro dzakarerekera.
Mune imwe nzvimbo yakafanana, mapofu emhando yekudzidzira data anozokanganisawo modhi yacho pachayo. Parizvino, mhando dzemitauro mikuru dzinotora mwedzi yakawanda kudzidziswa. Aya mamodheru zvakare anovimba nemadataset ane mashoma muhukuru. Ichi ndicho chikonzero ChatGPT ichingova neruzivo rushoma rwezviitiko zvakaitika yapfuura 2021.
mhedziso
Mienzaniso mikuru yemitauro ine mukana wekushandura zvechokwadi mabatiro atinoita tekinoroji nenyika yedu yese.
Huwandu hwedata huripo painternet hwakapa vaongorori nzira yekuenzanisira kuoma kwemutauro. Nekudaro, munzira, iyi mienzaniso yemitauro inoita kunge yatora manzwisisiro akaita seanhu enyika sezvairi.
Sezvo veruzhinji vanotanga kuvimba nemhando dzemitauro iyi kuti dzipe zvakabuda, vaongorori nevagadziri vave kutotsvaga nzira dzekuwedzera maguardrails kuitira kuti tekinoroji irambe ine hunhu.
Iwe unofunga kuti ramangwana reLLMs nderei?
Leave a Reply