Shieldstral Uyasebenza

I-Shieldstral iyimodeli yomthombo ovulekile engu-3B yamapharamitha yokuhlukanisa ukuphepha kokuqukethwe kwe-multimodal eyethulwe yi-Mistral AI, eyakhelwe phezu kwe-Ministral-3B. Imodeli ichaza kabusha ukuhlolwa kokuqukethwe kwezigaba ezimisiwe kwendabuko kube umsebenzi wemibuzo nezimpendulo ezimbili, futhi ivumela izinqubomgomo zokuhlola ukuthi zichazwe ngesikhathi sangempela ngemibuzo yolimi lwemvelo ngaphandle kokuqeqeshwa kabusha, ukuze ivumelane nezimo ezahlukene. I-F1 emaphakathi yemodeli ifinyelela ku-84.9% ebhentshimakhi yokuphepha kombhalo, kanti i-F1 yokuphepha kwe-multimodal ifinyelela ku-83.8%, zombili zisezingeni le-SOTA. Ingafakwa kusetshenziswa i-GPU eyodwa engu-16GB futhi isekela izilimi ezingu-12.

Ukuhlola okuvumelana nezinqubomgomo: Faka izinqubomgomo zokuphepha njengez પ્રશ્નo zolimi lwemvelo ukuze uzuze ukuhlola okwenziwe ngokwezifiso ngesikhathi sangempela ezimweni ezahlukene.
Ukutholwa okuhlanganisiwe kwe-multimodal: I-interface eyodwa icubungula ngesikhathi esisodwa ukuhlolwa kokuphepha kombhalo kuphela, izithombe, nokuqukethwe okuhlanganisa umbhalo nezithombe.
Ukuhlonza ukwephulwa kwemithetho ngezinga elinemininingwane: Isekela ukuhlukaniswa kokuphepha kusukela esinqumweni esiyisisekelo sika-yebo/cha kuya ezigabeni ezingama-73 ezinemininingwane.
Amaphuzu okuphepha alinganisiwe: Ikhipha amaphuzu okuphepha aqhubekayo ebangeni elingu-0–1 ngokusebenzisa i-softmax normalization, bese isebenzisa umkhawulo ongu-0.5 ukunquma ukwephulwa.
Ukufakwa okulula kumadivayisi asemaphethelweni: Ubukhulu bamapharamitha angu-3B buvumela ukucubungula kahle ku-NVIDIA GPU eyodwa engu-16GB, kunciphise izidingo zehadiwe.

Landela i-WeChat bese uphendula ngokuthi “Umthombo ovulekile” ukuze ujoyine iqembu lezingxoxo lamaphrojekthi e-AI omthombo ovulekile
Isakhiwo semibuzo nezimpendulo ezimbili: Guqula umsebenzi wokuhlola ube ukubikezela inani elinengqondo lamathokheni amabili okukhipha, u-“yes” no-“no”, bese uthola amaphuzu okuphepha aqhubekayo nge-softmax normalization.
Okokufaka okuhlelekile okunezinkambu ezintathu: Sebenzisa ifomethi evamile yezinkambu ezithi (isimo sokuhlola nezinga lokuqina), (umbuzo wokuthi kuphephile yini, yebo/cha), kanye (nokuqukethwe okuzohlolwa).
Ukuqeqeshwa kokuqhathanisa ngezinga elikhulu: Ngokusekelwe kumasampula ayizigidi ezingu-54.1 (izigidi ezingu-45.2 zemibhalo yomthombo ovulekile, izigidi ezingu-4.4 zemibhalo yokuqhathanisa eyenziwe ngokwenziwa, kanye nezigidi ezingu-4.5 ze-multimodal), imodeli iqeqeshwa ngamapheya okuqhathanisa anokuqukethwe okufanayo nemibuzo ehlukene ukuze ihlukanise izigaba ezifanayo.
Ukwengeza idatha eyenziwe ngokwenziwa: Sebenzisa i-LLM ukuze uguqule imibhalo ephephile ibe ukuhlukahluka okwephula imithetho, futhi wakhe ngokuzenzakalelayo amapheya emibuzo yokuqhathanisa phakathi kwesigaba okuhloswe kuso nezigaba ezihlobene, ukuqinisa ukuhlukanisa okunemininingwane.
Ukuhlelwa kahle kwe-LoRA + ukuhlanganiswa kwe-SLERP: Yenza i-LoRA fine-tuning esebenza kahle ku-Ministral-3B, bese uhlanganisa izindawo zokuhlola ezimbili ezihambisanayo eziqeqeshwe ngedatha yomphakathi nedatha eyenziwe ngokwenziwa usebenzisa i-spherical linear interpolation.

Ukulungiselela indawo: Faka imodeli ye-Shieldstral nohlaka lokucubungula ku-NVIDIA GPU eyodwa engu-16GB.
Chaza inqubomgomo yokuhlola: Bhala inkambu echaza isimo sokuhlola, isizinda somkhakha, kanye nezindinganiso zokuqina.
Bhala umbuzo wokuphepha: Yakha umbuzo wama-yebo/cha enkambini , isibonelo esithi “Ingabe okuqukethwe kukhuthaza udlame lomzimba?”.
Faka okuqukethwe okuzohlolwa: Faka umbhalo, isithombe, noma inhlanganisela yombhalo nesithombe enkambini bese ukuthumela kumodeli.
Thola isinqumo sokuphepha: Funda amaphuzu aqhubekayo alinganisiwe e-softmax akhishwe yimodeli, bese usetha umkhawulo ngokuya ngezidingo zebhizinisi ukuze wenze isinqumo esimbili sokwephulwa.

Ukuguquguquka kwezinqubomgomo: Izindinganiso zokuhlola zichazwa ngesikhathi sangempela ngolimi lwemvelo, ngaphandle kokuqeqesha kabusha imodeli ezimweni ezintsha.
Ukusebenza okuhamba phambili: Amapharamitha angu-3B afinyelela ezingeni le-SOTA kumabhentshimakhi okuphepha kombhalo nawe-multimodal, futhi aqhathaniseka namamodeli amakhulu ngokuphindwe ka-7.
Ukuhambisana nehadiwe: Isebenza nge-GPU eyodwa engu-16GB, yehlise kakhulu umgoqo wokufakwa kwangasese ezinkampanini.
Ukuhlanganiswa kwedatha: Ihlanganisa idatha ehlukahlukene engu-54.1M ngefomethi yomyalelo-umbuzo-idokhumenti, ihlanganise izilimi ezingu-12 nezimo eziningi.
Ukuchazeka kwezinqumo: Ikhipha amaphuzu aqhubekayo alinganisiwe esikhundleni samalebula okuhlukanisa angachazeki, okwenza kube lula kubasebenzi bokusebenza ukulungisa imingcele ngokuya ngezidingo.

Iwebhusayithi esemthethweni yephrojekthi: https://mistral.ai/news/shieldstral/
Inqolobane yamamodeli e-HuggingFace: https://huggingface.co/mistralai/Shieldstral-1.0-3B
Iphepha lobuchwepheshe le-arXiv: https://arxiv.org/pdf/2607.25857

Ukulawulwa kobungozi bokuqukethwe ezinkundleni zokuxhumana: Hlola ngesikhathi sangempela amaposi anombhalo nezithombe ashicilelwa abasebenzisi, futhi uvumelanise ngamandla nezindinganiso zokuhambisana zemiphakathi ehlukene.
Ukuvikelwa kokuphepha kwezingxoxo ze-AI: Thola ukuhlasela kokugqekeza emiyalweni yabasebenzisi nezimpendulo eziyingozi noma ezinokuchema ezikhishwa yimodeli.
Ukuhambisana kwezikhangiso nokumaketha: Hlola ngokuzenzakalelayo izakhi zombhalo nezithombe ezephula imithetho ezintweni zokukhangisa, uqinisekise ukuthi okuqukethwe okusatshalaliswa ezinkundleni ezahlukene kusemthethweni.
Ukuhlungwa kokuqukethwe kwemfundo eku-inthanethi: Thola ulwazi olungalungile ezintweni zokufunda nasekuxhumaneni, uvikele indawo yokufunda yabantwana.
Ukuhlolwa kokuphepha kwemibhalo yezinkampani: Hlola ngokuzenzakalelayo ulwazi olubucayi nokuqukethwe okwephula imithetho kumafayela ezilimi eziningi abiwe ngaphakathi enkampanini.