Chii chaitika
Google yakaburitsa autofinetune, sisitimu inoita otomatiki LLM mushure mekudzidziswa nekushandisa maAI maajenti kukwenenzvera nepamusoro mahyperparameter eSupervised Fine-Tuning (SFT) uye Reinforcement Kudzidza (RL). Ichi chishandiso chinobatanidza Google's Tunix raibhurari, Gemma modhi, uye Cloud TPUs, yakarongwa kuburikidza neAntigravity CLI uye Gemini Flash 3.7. Muzviitiko zvezvidzidzo, mumiririri akazvigadzirisa akazvigadzirisa senge LoRA mazinga uye mareti ekudzidza, kuvandudza kurongeka kwemuenzaniso uye zvibodzwa zvemubairo pasina kupindira kwemaoko.
Google yakazivisa kuburitswa kweautofinetune, chirongwa chakagadzirirwa otomatiki kudzidziswa kweMakuru Mutauro Models (LLMs). Iyo sisitimu inoshandisa yakazvimiririra yekutsvagisa loop apo mumiriri weAI anoongorora zvakare uye nekugonesa zvigadziriso zvekudzidzisa. Iyi nzira inofemerwa neyekutanga autoresearch purojekiti, iyo yakaratidza kuzvimiririra pre-kudzidziswa kuongorora.
Chishandiso chinokwirisa Google's yakazara AI stack, kunyanya kushandisa raibhurari yeTunix yekudzidziswa, maGemma modhi sehwaro, uye Cloud TPUs kuti compute. Iyo orchestration inobatwa neAntigravity CLI uye Gemini Flash 3.7. Chinangwa chikuru ndechekutsiva iyo yechinyakare yemanyorero kutenderera kwekugadzirisa hyperparameters ine otomatiki maitiro ayo anomhanyisa kuyedza kwehusiku humwe uye anoita kuvandudzwa kwakasimbiswa kuGit.
Muchidzidzo chekutanga, mumiririri akakwenenzvera google/functiongemma-270m-it modhi pagoogle/mobile-actions dataset vachishandisa Supervised Fine-Tuning (SFT). Iyo mumiriri inogadzirisa otomatiki paramita senge LoRA chinzvimbo, alpha, optimizer, uye chiyero chekudzidza. Mhedzisiro yacho yakaratidza kuvandudzwa kunoenderana mukukwanisa kwemuenzaniso kugadzira mafoni ekuita, zvichiratidza kugona kwemumiririri 'kukwira chikomo' kune chokwadi chiri nani.
Chidzidzo chechipiri chekuongorora chakanangana neKusimbisa Kudzidza (RL) uchishandisa nzira yeGRPO kudzidzisa Gemma 3 1B yekufunga kwemasvomhu padhata reGSM8K. RL inozivikanwa nekunzwa kwayo kune hyperparameters uye kusagadzikana. Iyo inozvimiririra mumiriri yakaratidza zvirinani zvigadziriso zveLoRA, tembiricha yekuburitsa, chirango cheKL, uye masisitimu ekusimudzira. Izvi zvakakonzera kukwenenzverwa kwe10% mumubairo wakakwana, zvichiratidza zvirinani nhamba uye fomati yechokwadi mumhinduro dzemuenzaniso.
Kwakabva mashoko: developers.googleblog.com ↗
Nei zvichikosha
Iyi budiriro inodzikisira zvakanyanya chipingamupinyi chekupinda kwemhando yepamusoro yeLLM-tuning nekubvisa kudiwa kwebhuku, kudzokorora kuyedza. Nekuita otomatiki kutsvaga kweakanyanya hyperparameter, inobvumira vanogadzira kuti vakwanise kuita zvirinani modhi nehunyanzvi hushoma uye nguva. Iyi shanduko yakanangana nekuzvimiririra yekutsvagisa zvishwe inogona kukurumidzira iyo iteration kutenderera kune vanogadzira AI, zvichiita kuti nzira dzepamusoro-soro dzekudzidzisa dziwanikwe uye dzinyatso shanda kune huwandu hwakawanda hwemasangano nevatsvaguri vega.
Autonomous inogadzirisa bhodhoro rakakosha mukuvandudza kweAI: nguva uye hunyanzvi hunodiwa kuti ugadzirise ma hyperparameters. Nekuita otomatiki maitiro aya, Google iri kuita kuti yemhando yepamusoro optimization iwanikwe kune vanogadzira vanogona kunge vasina hunyanzvi hwakadzama mukusimbisa kudzidza kana kunyatso-tuning mechanics.
Iko kubatanidzwa kweAI vamiririri mune yekudzidzira loop inomiririra shanduko yekuzvivandudza yega AI masisitimu. Kana vamiririri vachikwanisa kukwidziridza zvavo maparamendi ekudzidzisa, kumhanya kwemodhiyo kunogona kukurumidza, zvichigona kuderedza mutengo uye nguva ine chekuita nekugadzira maLLM akasarudzika emabasa chaiwo.
Ichi chishandiso chinonyanya kukosha kumasangano anoshandisa Google Cloud TPUs, sezvo ichipa yemuno, yakagadziridzwa mafambiro ekufambisa iyi hardware. Inotaridzawo basa riri kukura reagentic AI muinjiniya yesoftware uye yekutsvagisa workflows, ichifamba kupfuura yakapfava kodhi chizvarwa kune yakaoma yekuedza dhizaini uye kuuraya.
Interactive Mechanism: Iyo Inonyatsoshanda
Ongorora ari pasi tekinoroji kuseri kwekusimudzira uku uchipindirana.
An agent must create a draft calendar event for Tuesday at 2 p.m. Which evidence would establish the requested result?
Zvekutarisa zvinotevera
Tarisa kutorwa kweautofinetune munharaunda yekuvandudza uye chero inotevera inogadziridza kuraibhurari yeTunix. Tarisa uone mabhenji akazvimirira anoonesa mabhenji ekuita anonzi muGoogle nyaya dzezvidzidzo, kunyanya maererano nekugadzikana kwekuzvitonga kweRL tuning. Pamusoro pezvo, tarisa kana vamwe vakuru veAI vanopa vakaburitsa zvakafanana kuzvimiririra mushure mekudzidzisa maturusi, izvo zvinogona kuratidza shanduko yakakura yeindasitiri kuenda kune yekuzvigadzirisa modhi yekuvandudza mapaipi.
Kuzvimiririra kuoneswa kwemaitiro ekuita kwakakosha. Nepo Google ichishuma ~ 10% mubairo wekuvandudza mune RL nyaya yekudzidza, yechitatu-bato mabhenji anozodikanwa kusimbisa izvi mhedzisiro mumadhataseti akasiyana uye saizi yemhando.
Iko kugadzikana kwekuzvimiririra RL tuning inzvimbo yakakosha yekutarisa. RL haina kugadzikana, uye inoramba ichionekwa kuti mumiririri anobata sei nyaya dzemupendero kana kudzivirira kubirwa kwemubairo mune zvimwe zvakaomarara.
Adoption metrics eiyo autofinetune GitHub repository inoratidza kufarira kwemugadziri. Kana chishandiso chikawana kukosha kwakakosha, chinogona kukanganisa yakafara ecosystem yeLLM kudzidzisa maraibhurari uye maturusi.