I-Mistral AI imemezele izolo (4 Agasti) ukwethulwa kwemodeli ye-AI yokuhlola okuqukethwe i-Shieldstral, enamapharamitha angu-3B (3 billion), enezisindo ezivulekile futhi ekhishwa ngaphansi kwelayisensi ye-Apache 2.0.
Imodeli isiyatholakala epulatifomu i-Hugging Face, isekela izilimi ezingu-12 futhi ingasebenza ku-GPU eyodwa engu-16GB. I-Mistral ithi le modeli inekhono lokuphepha kokuqukethwe elilingana namamodeli avulekile amakhulu ngokuphindwe ka-7, futhi izuze i-SOTA emkhakheni wokuhlola okuqukethwe okunezindlela eziningi.
Qaphela: I-SOTA isifinyezo se-state-of-the-art, esisho iqoqo lamamodeli noma ama-algorithm asezingeni eliphezulu kakhulu futhi asebenza kahle kakhulu emsebenzini othile we-AI noma ekuhlolweni kwebenchmark. Leli zinga lihlala lishintsha njengoba ubuchwepheshe buthuthuka futhi kushicilelwa ucwaningo olusha.
I-Mistral iveze ukuthi iningi lamamodeli okuvikela lifaka uhlelo lwezigaba zokulimaza ngaphakathi kwezisindo zemodeli. Uma umkhiqizo usetshenziswa kwesinye isimo, abathuthukisi ngokuvamile kudingeka baqeqeshe kabusha imodeli.
I-Shieldstral yona ifaka izinqubomgomo zokuhlola kokuqukethwe kokuqukethwe okokufaka: osebenzisa yona angabhala umbuzo othi “yebo noma cha”, bese enikeza isimo sokuhlola kanye nencazelo yezinga lokuqina. Imodeli ibe isikhiqiza isikolo sokuphepha esilinganisiwe kusuka ku-token eyodwa yokukhipha.
Imodeli iguqula umsebenzi ngamunye wokuhlola ube yimibuzo nezimpendulo ezimbili, lapho okokufaka kuqukethe izinkambu ezingu-3 ezinelebula:
<Instruct>: Ichaza isimo sokuhlola nezinga lokuqina.
<Query>: Ibuza umbuzo owodwa othi “yebo noma cha”, isibonelo, “Ingabe lokhu okuqukethwe kukhuthaza udlame lomzimba?”
<Document>: Inikeza okuqukethwe okuzohlolwa, okungaba i-prompt, impendulo, inhlanganisela ye-prompt nempendulo, noma isithombe esinombhalo ongakhethwa.
Ngesikhathi sokwenza inference, imodeli ifunda kuphela amanani anengqondo ama-token amabili athi “yebo” no-“cha”, bese iwaguqula abe isikolo esiqhubekayo ngokusebenzisa i-softmax normalization, iphinde ikhiphe isinqumo esimbaxambili isebenzisa u-0.5 njengomkhawulo. Le ndlela ingasetshenziswa ekuhlukaniseni ama-prompt, ekuhloleni izimpendulo, ekutholeni ukwenqaba ukuphendula nasekuhloleni ubuthi bokuqukethwe.
