Vatsvakurudzi vanobva kuYunivhesiti yeNijmegen (Netherlands) vakagadzirira kuvhurika kwemazana makumi mana emitauro mikuru yemitauro uye 40 modhi yekugadzira mifananidzo kubva kune tsananguro yezvinyorwa, izvo zvinoziviswa nevagadziri sekuvhurika. Nekuda kwekuti maitiro ekuvhurwa kwemamodhi ekudzidza muchina achiri kuumbwa, mamiriro ezvinhu amuka apo, pasi pechifukidziro chekuvhurika, mamodheru akagovaniswa ane rezinesi inodzikamisa chiyero chekushandiswa (semuenzaniso, mazhinji. mhando dzinorambidza kushandiswa muzvirongwa zvekutengesa). Zvakare, kazhinji vagadziri havapi mukana kune iyo data inoshandiswa mukudzidziswa, havaburitse ruzivo rwekuita, kana kusavhura zvizere kodhi inoperekedza.
Mamodheru mazhinji anotengeswa se "akavhurika" anofanirwa kutorwa se "zviyero zvakavhurika" kana zvakanyanya "zviyero zvinogoneka" nekuti zvinouya pasi pemarezinesi anorambidza kushandiswa mune zvekutengesa. Vatsvakurudzi vekunze vanogona kuedza nemhando dzakafanana, asi havakwanisi kugadzirisa modhi kune zvavanoda kana kuongorora kushandiswa. Inopfuura hafu yemamodheru haapi ruzivo rwakadzama nezve data rinoshandiswa pakudzidziswa, uye haabudise ruzivo nezve mukati medhizaini uye dhizaini.
Mamodheru akavhurika zvakanyanya ndeaya BloomZ, AmberChat, OLMo, Vhura Mubatsiri uye Yakagadzikana Diffusion, ayo anoburitswa pasi pemarezinesi akavhurika pamwe nekwakabva data, kodhi uye API kuita. Mienzaniso kubva kuGoogle (Gemma 7B), Microsoft (Orca 2) uye Meta (Llama 3), yakamisikidzwa nevagadziri seyakavhurika, yaive pedyo nekupera kweiyo chinzvimbo, sezvo vasingapi mukana wekuwana data, usaburitsa ruzivo rwehunyanzvi. yekushandiswa, uye uremu coefficients modhi inogoverwa pasi pemarezinesi anodzikamisa chiyero chekushandisa. Iyo yakakurumbira Mistral 7B modhi yaive pakati pechiyero, sezvo ichipihwa pasi perezinesi rakavhurika, asi inongonyorwa zvishoma, haiburitse data rinoshandiswa mukudzidziswa, uye haina kodhi yakazara yakazaruka inoperekedza.
Vatsvakurudzi vakakurudzira nzira ye14 yekuvhurika kwemhando dzeAI, inovhara kugoverwa kwekodhi, data yekudzidzira, uremu, kusiyana kwedata uye coefficients yakagadziridzwa kuburikidza nekudzidza kwekusimbisa (RL), pamwe nekuwanikwa kwekugadzirira-kushandisa-pakeji, APIs, magwaro uye tsananguro dzakadzama kuita.
Zvinoenderana netsanangudzo yedhizaini yeAI yakavhurika yakatsanangurwa neOSI (Open Source Initiative), maitiro makuru ekuvhurika kweAI sisitimu kupihwa kwemikana yekushandisa chero chinangwa pasina chikonzero chekuwana mvumo yakaparadzana; kudzidza kushanda kwegadziriro uye kuongorora zvikamu zvayo; kuita shanduko kune chero chinangwa; endesa kune vamwe vanhu vezvese zviri zviviri shanduro yekutanga uye edition mushure mekuchinja kwaitwa.
Kugonesa shanduko kuti dziitwe, iyo AI system inofanirwa kusanganisira:
- Ruzivo rwakadzama nezve data rinoshandiswa mukudzidziswa uye nzira yekudzidzisa. Panofanirwa kuve neruzivo rwakakwana kuti nyanzvi yekuvandudza ikwanise kugadzira zvakare yakaenzana AI sisitimu ari ega, achishandisa imwechete kana yakafanana data yekudzidziswa.
- Kuvepo kweiyo source code iyo inokutendera kuti mese mutange iyo AI system uye kuita maitiro ekuidzidzisa (mutafura yakurukurwa pamusoro, mu "code" column yemhando dzakawanda "~" inoratidzwa, izvo zvinoreva kuwanikwa kwechikamu. kodhi kana kodhi iripo yekumhanyisa modhi, asi hapana kodhi yekudzidzisa kana kugadzira modhi). Iyo kodhi inofanirawo kuvhara nzvimbo dzakadai sepreprocessing, kusimbiswa kwedata, uye tokenization. Mukuwedzera, tsananguro yakadzama yemhando yekuvaka inofanirwa kupihwa.
- Model parameters (weighting coefficients), zvichireva kuvapo kweyakagadzirira-kushandisa-nyika chidimbu mushure mekudzidziswa kana kuvepo kwekupedzisira optimized shanduro yemuenzaniso.
Source: opennet.ru
