- I-SmolVLM-256M: Imodeli yolimi lombono wamapharamitha ayizigidi ezingu-256, elungiselelwe amadivayisi alinganiselwe kwezinsizakusebenza kanye nokuphatheka kalula.
- Ukwakhiwa kwesakhiwo nge-patch-16/512 esekelwe ku-SigLIP encoder (amapharamitha angu-93M) kanye nokucindezelwa kwethokheni ukuze kube nesisombululo esiphezulu kanye nokuzinza ngesikhathi sokuqeqeshwa.
- Amakhono e-multimodal: izincazelo zesithombe, imibuzo yamadokhumenti, kanye nokuhlaziywa kwegrafu, okuwusizo kwezemfundo kanye namabhizinisi anamandla aphansi.
I-SmolVLM-256M ivele endaweni yobuhlakani bokwenziwa njengemodeli yolimi lombono oluncane kakhulu (i-VLM) kuze kube manje. Yasungulwa yi -Hugging Face , lobu buchwepheshe buhlose ukusebenza kahle kakhulu futhi kufinyeleleke kalula, ngisho nakumadivayisi anezinsizakusebenza zekhompyutha ezilinganiselwe. Yakhelwe ngokuphatheka nokusebenza engqondweni, le modeli ithembisa ukuguqula indlela esisebenzisana ngayo ne-AI, kokubili kumadivayisi omuntu siqu kanye nezinhlelo zokusebenza zebhizinisi.
Enye yezinto ezikhangayo kakhulu ze-SmolVLM-256M ubuncane bayo. Njengoba inamapharamitha ayizigidi ezingu-256 kuphela , le modeli iyakwazi ukwenza imisebenzi eyinkimbinkimbi njengokukhiqiza izincazelo zezithombe, ukuhlaziya amavidiyo amafushane, nokuphendula imibuzo mayelana namadokhumenti e-PDF. Le ndlela ayigcini nje ngokuthuthukisa ukusetshenziswa kwehadiwe kodwa futhi ivumela ukuthi isetshenziswe kumadivayisi alula njengama- laptops ane-RAM engaphansi kwe-1GB.
Izici zobuchwepheshe ezigqanyisiwe
Isisekelo sempumelelo ye-SmolVLM sisekwakhekeni kwayo okuthuthukisiwe. Isebenzisa i-visual encoder ebizwa ngokuthi i-SigLIP base patch-16/512 , enamapharamitha ayizigidi ezingu-93 . Le encoder ayincane nje kuphela kuneyandulelayo yamapharamitha ayizigidi ezingu-400 , kodwa futhi ithuthukisa ukuxazululwa kwezithombe ezicutshunguliwe. Lolu shintsho luphefumulelwe ucwaningo lwangaphambilini oluvela ku -Apple kanye ne-Google , okubonisa ukuthi ukuxazululwa okuphezulu kokubonwayo kungathuthukisa kakhulu ukuqonda ngaphandle kokukhulisa usayizi wemodeli.
Ngaphezu kwalokho, i-SmolVLM isebenzisa amasu okucindezela amathokheni athuthukile, okuvumela ukumelwa kwesithombe okusebenzayo kakhudlwana. Isibonelo, abahlukanisi bezithombe ezingaphansi manje bamelelwa ithokheni elilodwa, esikhundleni samathokheni amaningi, okuye kwaba negalelo ekuzinzeni okukhulu kanye nekhwalithi ethuthukisiwe ngesikhathi sokuqeqeshwa kwemodeli.
Amakhono we-Multimodal
Phakathi kokusebenza kwe-SmolVLM kunemisebenzi efana nale:
- Izincazelo zesithombe: Ilungele izinhlelo zokusebenza ezidinga isingeniso esibonakalayo esinemininingwane, njengamathuluzi okufundisa noma i-ecommerce.
- Izimpendulo zemibuzo emayelana namadokhumenti: Kusukela kumadokhumenti e-PDF kuya embhalweni oskeniwe, imodeli ihlonza futhi ihlaziye okuqukethwe okubukwayo nombhalo.
- Ukuhlaziywa kwamagrafu nemidwebo: Isixazululo esibalulekile sezinkampani ezisebenza ngedatha ebonakalayo eyinkimbinkimbi.
Lezi zici zenza le modeli ifaneleke kakhulu kumaphrojekthi afana nokwenza kahle kwamadokhumenti nokucabanga okuyisisekelo okubonakalayo, ikakhulukazi ezindaweni zemfundo nasezindaweni zebhizinisi.
Ukusetshenziswa okusebenzayo kanye nokwenza kahle
I-Hugging Face yethule i-SmolVLM ngezinjongo zokuthuthukisa izindleko . Isibonelo, izinkampani ezisebenza ngenani elikhulu ledatha ebonakalayo zingazuza ekusetshenzisweni okuphansi kwezinsizakusebenza zemodeli . Ukucubungula izithombe ezifika ku- 1 million ngenyanga nge-SmolVLM kungabangela ukonga okukhulu uma kuqhathaniswa namamodeli amakhulu, avamile.
Ngaphezu kwalokho, izinkampani ezifana ne -IBM sezivele zihlanganise lo modeli ezinhlelweni zokusebenza ezifana ne- Docling , isofthiwe yokucubungula amadokhumenti. Umphumela uba ukusebenza kahle okukhulu ekuphathweni kwedatha, izindleko zokusebenza ezincishisiwe, kanye nokuncintisana okwandisiwe emakethe.
Ukuqeqeshwa kanye nedatha esetshenzisiwe
Imodeli yaqeqeshwa kusetshenziswa amasethi edatha amabili ayinhloko: i-Cauldron ne -Docmatix . I-Cauldron ifaka amasethi edatha ekhwalithi ephezulu angaphezu kwama- 50 ahlanganisa izithombe nombhalo, kuyilapho i-Docmatix igxile kumadokhumenti askeniwe kanye namagama-ncazo awo. Le ndlela ivumele imodeli ukuthi ilungiselelwe imisebenzi ethile njengokuhlaziywa kwedokhumenti kanye nencazelo yesithombe.
Kuye kwabekwa phambili nemisebenzi efana nokuqonda imidwebo yesayensi nokuhlaziya izibalo eziyisisekelo. Nakuba ukusebenza kwayo kuvelele emisebenzini eminingi, kubalulekile ukuqaphela ukuthi amamodeli amakhulu asasebenza kangcono kuneSmolVLM ezinkingeni zokucabanga ezithuthukisiwe.
Imikhawulo nezinselele
Naphezu kwezinzuzo zayo eziningi, i-SmolVLM ayiyona engenamingcele . Izifundo zakamuva zibonise ukuthi amamodeli amancane, njengalena, avame ukuba nenkinga yokucabanga okunengqondo okuyinkimbinkimbi. Lokhu kungenxa yokuthi, nakuba eqaphela amaphethini angaphandle kudatha, avame ukwehluleka ukusebenzisa lolo lwazi ezimweni ezintsha.
Ngaphezu kwalokho, nakuba ukusetshenziswa kwayo kwamandla okuphansi kwezingxenyekazi zekhompiyutha kungamandla, akukwenzi ukuthi ifanelekele izimo eziyinkimbinkimbi kakhulu lapho kudingeka khona ukucutshungulwa okunemininingwane noma okucashile, njengokucwaninga okuthuthukile noma izinhlelo zokusebenza ezithile zesayensi.
I-SmolVLM-256M imele isisombululo esisha nesifinyelelekayo salabo abafuna ukusebenzisa i-AI kumadivayisi avinjelwe izinsiza. Inhlanganisela yayo yamakhono e-multimodal, ukusebenza kahle kwekhompyutha, kanye nokuguquguquka kuyenza ibe inketho ekhangayo kubo bobabili abathuthukisi namabhizinisi. Ngalezi ntuthuko, i-Hugging Face ibonisa ukuthi ikusasa lobuhlakani bokwenziwa alikho kuphela kumamodeli amakhulu futhi ayinkimbinkimbi, kodwa futhi ekwakhiweni kwezakhiwo ezincane nezisebenzayo ezenza intando yeningi ukufinyelela kulobu buchwepheshe.
