If you’re looking for exact figures, I’d recommend checking their official documentation or recent benchmarks, as these metrics can change with updates and optimizations. Would you like to know more about its architecture or applications?

Imwe yemashandisirwo ari nyore emhando yemutauro muhombe (LLM) yakaita sedhatabhesi reruzivo. MaLLM akadzidziswa pane yakakura dataset yeruzivo rwakapfuma, iyo inopindirana seChatGPT yaita kuti zvive nyore kutora. Kana iwe uchikurudzira ChatGPT kuti idzose zvemukati kubva mubhaisikopo kana bhuku, semuenzaniso, unenge uchingoshandisa iyo modhi kugona kurangarira ruzivo rwayakaratidzwa panguva yekudzidziswa kwayo. Asi zvakadini kana isina kudzidziswa pane Terminator script, kana kana uremu hwayo husingapi kukosha kumabasa aDickens? Kuti tipe mibairo chaiyo uye yakakosha kune chero yakapfava yekushandiswa kwenyaya, sekudzoreredza ruzivo rwekutanga, LLM dzinoda yakaomesesa indexing uye nzira dzekutora dzinokwanisa kuwana huwandu hwakawanda hweruzivo nemazvo.

Kunzwisisa kugadzirwa kweLLM uye kudzidziswa

Zvemukati zveLLM zvinogadzirwa kuburikidza nemaitiro anozivikanwa sekufembera kwechiratidzo chinotevera , chinova nechokwadi chekuti mhinduro dzakafanira maererano nemamiriro ezvinhu, akasiyana, uye anoratidza kunzwisisa sekunge munhu. Heano mashandiro anotevera token kufanotaura kunoshanda, nhanho nhanho:

1. Input Processing: Paunonyora kukurumidza kana mubvunzo, iyo yekuisa inoshandurwa kuita tokeni: mazwi kana zvidimbu zvemashoko.

2. Kunzwisisa Mamiriro ezvinhu: Muenzanisi unotarisa zviratidzo zvawakaupa uye, zvichibva pakudzidziswa kwawakaita, unoedza kunzwisisa mamiriro ezvinhu, izvo zvinosanganisira zvose kubva kumusoro uri kutaurwa kusvika kune toni yaungave uri kushandisa.

3. Inotevera Chiratidzo Kufanotaura: Uchishandisa mamiriro ayo anonzwisiswa, modhi yacho inofanotaura kuti chii chingangove chinotevera chiratidzo. Hakusi kungofembera zvichibva pashoko rapfuura; iri kufunga nezvese mamiriro ehurukuro kusvika panguva iyoyo.

4. Tokeni Sarudzo: Kana yangofanotaura huwandu hweanotevera tokeni, inosarudza imwe. Sarudzo iyi yakavakirwa pamukana-chiratidzo chinonyanya kuuya chinotevera zvichienderana nedata rakadzidziswa modhi. Izvo zvakakosha kuti ticherechedze, zvisinei, kuti pane kumwe kurongeka pano zvakare, izvo zvinobatsira kuburitsa akawanda akasiyana uye echisikigo-anonzwika mhinduro.

5. Output Generation: Chiratidzo chakasarudzwa chinozoshandurwa zvakare kuita mavara anoverengwa nevanhu. Kana mhinduro isina kukwana (iyo kazhinji isiri mushure mechiratidzo chimwe chete), maitiro anodzokorora. Chiratidzo chitsva chinowedzerwa kune kutevedzana, uye modhi inofanotaura chiratidzo chinotevera zvichibva pane ino yakagadziridzwa mamiriro.

6. Iterative Refinement: Iyi nzira yekufanotaura chinotevera tokeni uye kuiwedzera kune iyo kutevedzana inodzokororwa kusvikira modhi yasvika painomira. Izvi zvinogona kuva apo mhinduro inosvika pahurefu hwakati, muenzaniso unofanotaura chiratidzo chinoreva kupera kwemutsara kana ndima, kana kuti inozadzisa mirairo yakaiswa mukukurumidza.

Miganho yekumanikidza muLLM kudzidziswa

Kana LLM ichifanotaura chiratidzo, iri kunyatso tora uye kushandisa iyo yakadzvanywa ruzivo rwakaiswa mukati mehuremu hwayo kuti ibudise zvinoenderana nemamiriro ezvinhu. Nenzira iyi, LLM kudzidzisa magirazi kudzvanya database. Sezvinongoita dhatabhesi yakagadziridzwa kuyeuka kazhinji inowanikwa data nekukurumidza, iyo LLM yakagadzirirwa kudzoreredza ruzivo-chaiyo yakapindirwa ndangariro-kubva kune huremu hwayo. Kugona uku kunoibvumira kuburitsa mhinduro chaidzo kumibvunzo pamusoro pezvinhu zvinozivikanwa zvayakasangana nazvo panguva yekudzidziswa kwayo, sekubvunza dhatabhesi kuti uwane ruzivo rwakanyatsotsanangurwa. Nekudaro, zvipingamupinyi zvinomuka kana modhi ikasangana zvisingazivikanwe kana kuti zvakavanzika zvemukati. Semuyenzaniso, paunobvunza LLM kuti kune mavhesi emuBhaibheri, inoatora izwi neshoko, asi haigone kutora izwi neshoko chero pfungwa yayasina "kupupurira" zvakanyanya panguva yekudzidziswa, sezvo huremu hunoenderana neiyo pfungwa zvakare. kusakosha. Nenzira iyo zvakare, iyo LLM yakafanana nedatabase. Sezvinongoita dhatabhesi rinogona kungodzosera dhata rakachengetwa zviri pachena mukati maro, iyo LLM inogona kunetsekana nekugadzira zvirimo pamisoro yayasina kuona zvakanyanya panguva yekudzidziswa.

Ehe, maLLM ari pamusoro pechikamu chekuenzanisa uku, sezvo vaine modhi yepasirese mukati inovabvumira "kunzwisisa" zvinhu kunze kwekungotarisa chete. Nekudaro, kuwedzeredza uku kunotibatsira kunzwisisa zvimwe zvipimo zvakakosha munzira iyo maLLM anodzidziswa kugadzira zvirimo.

Zvimwe zvinogumira pakudzidziswa kweLLM

Uyezve, iyo inotevera tokeni yekufungidzira system ine zvimwe zvipingamupinyi zvemukati zvinobva munzira yayo yakakosha yekugadzira zvinyorwa:

• Context Window Size: Chimwe chezvipingamupinyi chikuru chemuenzaniso wewindow saizi - kuwanda kwehuwandu hwemavara (mumatokeni) iyo modhi inogona kutariswa kana uchifanotaura. Pamienzaniso yakawanda, kusanganisira shanduro dzekare dzeGPT, hwindo iri harina kukura zvakakwana kuchengetedza mamiriro pamusoro pehurukuro refu kana zvinyorwa, izvo zvinogona kutungamirira mukurasikirwa kwekubatanidzwa mumagwaro marefu kana nhaurirano dzakaoma dzinoda kuchengetedza mamiriro ezvinhu kunze kwezviratidzo zvakapfuura.

• Generalization vs. Kunyatsojeka: Kunyange iwo mamodheru akadzidziswa pamaseti akakura, kugona kwavo kuita zvakazara kubva padzidziso iyi dzimwe nguva kunogona kutungamirira kuti vabudise zvinyorwa zvegeneric kana zvisina kujeka. Vanogona kupotsa mucherechedzo mukugadzira mhinduro dzakanangana kana dzakanangana dzinoda nzwisiso yakadzama kana ruzivo rwechizvino-zvino kunze kwedata ravo rekudzidziswa.

• Kushaikwa KweKunze Ruzivo Ruzivo: Inotevera tokeni yekufembera modhi inogumira kune iyo ruzivo rwuri mukati mekudzidzisa kwavo dataset. Ivo havagone kuwana kana kubatanidza ruzivo rutsva mushure mekudzidziswa, izvo zvinoreva kuti vanogona kukurumidza kuita zvechinyakare kana kushaya mamiriro azvino, senge zvichangobva kuitika, zvakawanikwa, kana nyaya dzirikuitika.

• Kudzokorodza uye Kufanotaura: Iyo algorithmic chimiro chekufanotaura kwechiratidzo chinotevera dzimwe nguva chinogona kukonzera kudzokorora kana kufanofungidzira kugadzirwa kwemavara. Sezvo modhi yacho ichiwanzofarira ma tokens ayo anogona kutevedzwa nenhamba zvichienderana nemamiriro ezvinhu, anogona kuwira muzvishwe kana kusarudza zvirevo zvakajairika, zvichideredza kusiyanisa kwezvinobuda.

Retrieval augmented generation (RAG) yakatsanangura

Sezvambotaurwa, maLLM anoburitsa mhinduro zvichibva pane huremu hwavakapa kune akasiyana e data panguva yekudzidziswa. Aya huremu anoratidza kukosha kana kukosha kwakasiyana-siyana kwedata rekuisa inonzwisiswa nemuenzaniso. Kana kukurumidza kwemushandisi kuchisanganisira zvinhu zvisina kumiririrwa zvakanyanya mudhata rekudzidziswa, modhi inogona kutadza kuburitsa mhinduro chaiyo kana yakakodzera.

Kana nhaurirano ikadarika hwindo remukati meLLM, kana kukurumidza kukadarika muganho wehuremu hwakakosha mune yeLLM yekudzidzisa dhataset (zvichireva kuti haigone kurangarira chaizvo mhinduro iri kutsvagwa nemushandisi), modhi yacho inowanzotsamira pane yekunze vector yekutsvaga database. , iyo inoibvumira kutsvaga mamiriro akakodzera kana data nyowani inogona kuwedzerwa kune kukurumidza kubva kumushandisi. Iyi nzira inozivikanwa sekudzoreredza augmented generation (RAG).

"Vector kutsvaga kubudirira"

Iyo RAG maitiro anoitwa kuti agoneke kuburikidza nevector yekutsvaga dhatabhesi: yemhando yepamusoro yedatabase inochengeta uye inogadzirisa data semavheji. Aya mavheji anomiririra data munzvimbo yakakwira-dimensional, iyo imwe neimwe dimension inobata chimwe chikamu chezvinoreva data, zvichibvumira kumiririrwa kwehukama hwakaoma uye hunhu. Muchirevo chechinyorwa uye mutauro, vector yekutsvaga dhatabhesi inoshandisa matekiniki akadai sekumisikidza kushandura zvinyorwa kuita manhamba vectors. Shanduko iyi inogonesa sisitimu kuyera kufanana kwesemantic pakati pezvidimbu zvakasiyana zvechinyorwa nekuverenga nhambwe dziri pakati pemavekita anoenderana munzvimbo ino ine mativi mazhinji.

Munguva yeRAG, zvese mubvunzo (kureva, kupinza kwemushandisi kuLLM) uye data rakachengetwa (senge zvinyorwa, zvinyorwa, kana mitsara) zvinoshandurwa kuita mavekita vachishandisa mavara akaiswa. Izvi zvakamisikidzwa zvinoshandura dhata rezvinyorwa kuita manhamba mavheji uko zvirevo zvakafanana zvinomisikidzwa kune mapoinzi ari pedyo munzvimbo yevector. Dhatabhesi rinobva raverengera nhambwe dziri pakati pemubvunzo vector uye mavekita e data rakachengetwa kuti rione kuti zvirevo zvezvinyorwa zvine hukama hwakadii. Iyo dhatabhesi inotora mapoinzi edhata (zvinyorwa zvemukati) ane mavekita ari padyo nemubvunzo vector, kureva, iwo ane semantically akafanana zvakanyanya nekuisa. Aya mapoinzi edata anoonekwa se "vavakidzani vepedyo" maererano nemamiriro uye zvinoreva.

Vavakidzani vepedyo ava vanopa ruzivo rwakakodzera, rwekuwedzera iyo iyo LLM inogona kunge isina kuwana mukati meiyo data yekudzidziswa, iyo inogona zvakanyanya kunatsiridza huroyi, kukosha, hupfumi, uye akasiyana ezvinobuda muLLM. Sam Altman, pakati pevamwe, akatsigira nzira ye "vector yekutsvaga kubudirira"-kuvimba neRAG yekugadzira vamiririri, pane kuenzanisira kugadzirisa chete.

RAG seimwe nzira yekugadzirisa zvakanaka

Kunyatsogadzirisa LLM kunosanganisira kugadzirisa huremu hwemodhi zvichibva pakuwedzera kudzidziswa pane yakatarwa dataset kuti uwedzere kuita kwemamwe mabasa kana kunatsiridza kunzwisisa mune mamwe madomasi. Haisi chete maitiro aya anononoka pane kumhanya kwehunyanzvi, zvichireva kuti mamodheru akakwenenzverwa anopera basa nekukasira sezvaanogadziridzwa, zvakare haigadzirise nyaya yedata nyowani.

Mukupesana, RAG inogonesa iyo modhi kuti iwane ekunze dhatabhesi munguva chaiyo kuti itore iyo yazvino ruzivo rwakakodzera kumubvunzo uri pedyo. Kunyangwe iyo yepasi modhi isati yavandudzwa kana kukwenenzverwa nguva pfupi yadarika, inogona kuburitsa mhinduro dzinosanganisira data razvino. Mienzaniso inoramba yakakosha kwenguva yakareba nokuti inogona kuenderana nedheta itsva uye kuchinja mamiriro ezvinhu kuburikidza nekudzorerwa kwezvinyorwa zvemashoko ekunze.

RAG inonyatso bhiridha mukaha uripo pakati pekudzidza kwakadzama nemaitiro echinyakare ekutoresa ruzivo. Nekuita izvi, zvinosimudzira masimba ezvese — kudzidza kwakadzama kunzwisisa kwemamiriro ezvinhu kune simba uye nemazvo ekutsvagisa ruzivo. Iyi nzira yakasanganiswa inobvumira maLLM kuti abudise mhinduro dzakadzama, dzakadzama, uye dzakapfuma.

Kugadzirisa zvimwe zvipimo zveLLMs

Pamusoro pekugadzirisa zvakanaka, RAG zvakare inogadzirisa matambudziko akambocherechedzwa ane hukama neyakajairwa LLMs:

• Kuwedzera Kunzwisisa Kwemamiriro ekunze: RAG inotambanudza hwindo remamiriro echinyakare LLMs nekutora-kusvika-zuva kana ruzivo rwakadzama runosimudzira mhinduro dzemodhi.

• Kuvandudza Hunhu uye Huroi: Panzvimbo pekuvimba chete nemapateni akadzidzwa panguva yekudzidziswa, RAG inobvumira modhi kupinza ruzivo rwakadzama kubva mumagwaro akadzoserwa mumhinduro dzayo, ichiita kuti dzisangova dzakanyanya kunaka asiwo dzakanangana nemubvunzo uri pedyo.

• Kuderedza Kudzokorora uye Kufanotaura: Nekudhonza zvine simba seti dzakasiyana dzeruzivo kumubvunzo wega wega, RAG inogona kusiyanisa mhinduro dzemuenzaniso zvakanyanya. Kusiyana uku kunobatsira mukudzikisira kudzokorora uye kufanotaura kunowanzoonekwa mumamodheru akachena, sezvo data rekunze rinounza zvirevo zvitsva uye ruzivo muhurukuro.

Matambudziko uye inodiwa shanduko yeRAG

RAG inouya nematambudziko ayo, zvisinei-kureva latency uye kushaya njere. Funga nezve kutendeuka-based agent chatbot hurukuro apo mushandisi anotumira kukurumidza, iyo LLM inopfira matinji mashoma anoratidza kuti inoda mamwe mamiriro, vector yekutsvaga dhatabhesi inotora iri padyo-yemuvakidzani mamiriro kuburikidza nemushandisi wekuisa kukurumidza, uye ipapo zvese zvinotumirwa. kuLLM zvakare kuti uwane inference. Zvadaro, inguva yemushandisi yekupindura, zvichingodaro.

Mune ino sisitimu, yega yega mushandisi kukurumidza anotanga akawanda-nhanho oparesheni apo nhanho imwe neimwe inowedzera kune yakazara nguva yekugadzirisa. Iko kumhanya kwemaitiro ese kunoenderana nekuti nekukurumidza sei iyo vector yekutsvaga dhatabhesi inogona kutora iyo inodiwa mamiriro. Kana mubvunzo wepa database wakaoma kana kuti dhatabhesi pachayo yakakura uye isina kurongeka zvakakwana, kutora uku kunogona kuunza kunonoka kukuru. Pamusoro pezvo, kunyanya munhaurirano dzakaoma kunzwisisa, kutevedzana uku kwechizvarwa uye kudzoreredza kungangoda kudzokororwa kakawanda kuti kunatsiridza mhinduro zvakakwana. Iyi iterative denderedzwa inogona kusanganisa iyo latency, zvichitungamira kune kunonoka kupindirana pane zvingave zvichigoneka neyakangoita generative modhi inotsamira chete pane data remukati.

Uyezve, hungwaru hweRAG-yakapfuma LLM hunoenderana zvakanyanya nemhando uye kukosha kweruzivo rwakatorwa kubva kune vector yekutsvaga database. Kana zvirimo mudhatabhesi zvisina kuzara, zvechizvino-zvino, kana kuchengetedzwa zvakanaka, kushandiswa kweruzivo rwakadzoserwa kunogona kuve kushoma, zvichikanganisa hungwaru hwese hwemhinduro.

Kunyangwe kana data rekunze remhando yepamusoro rakadzoserwa, dambudziko rinoramba riri pakuti ruzivo urwu rungabatanidzwa sei muhurongwa hwemhinduro huripo hweLLM. Iyo modhi haifanire kungobatanidza iyi data rekunze chete asi zviite nenzira yakafanira uye inowirirana. Kusarongeka pakati pekudzidziswa kwemuenzaniso uye chimiro che data rekunze zvinogona kutungamira kune mhinduro dzakanyatsoenderana nehunyanzvi asi dzisina kubatana.

Chizvarwa chinotevera cheLLMs

Chizvarwa chinotevera cheLLMs chingangobatanidza vector yekutsvaga-based RAG uye tsika dzechinyakare kudzidziswa/kunatsa-tuning nzira pamwe chete, pamwe nehurongwa hwekugadzirisa data (semuenzaniso SQL dhatabhesi yeTradFi musika data uye inoenderana nezvemari nhau). Pfungwa yekuve nemupi weLLM 'pano' uye yakaparadzana vector yekutsvaga dhatabhesi 'pamusoro apo' ichabatana kuburikidza nemhando nyowani iyo intuitively inowedzera yavo indexed yekushanda memory kune emunharaunda maSSD ane terabytes eiyo vectorized mamiriro.

Nzvimbo uye Nguva yakatopa Humbowo hweSQL-uchapupu hweZK hunosimbisa huchokwadi uye kutapudza kweSQL database processing-kune vatengi uye ichangobva kutumirwa Uchapupu hweVector Search, iyo inoita zvakafanana kune vector yekutsvaga kutsvaga. Idzi humbowo humbowo hunovhura nzira yeramangwana apo maLLM anogona kubatanidza mamiriro matsva, kuwana yakafaranuka uye yakawedzera nuanced spectrum yedata munguva chaiyo, uye kubatanidza yakarongeka magadzirirwo edata kuti abudise analytics ane hungwaru, zvese nenzira inoteedzeka, inogoneka. Kufambira mberi uku kunozopedzisira kwawedzera huwandu hwezvikumbiro zveLLMs, vachiwedzera kushandiswa kwavo muzvikamu zvinotsamira zvakanyanya pane-kusvika-paminiti data, senge zvemari masevhisi, kuunganidzwa kwenhau, uye kuongororwa kwenjodzi, nekudaro kufambisa mberi kunotevera kweAI. -inofambiswa innovation.