I-Fastino Labs Open-Sources GLiGuard: Imodeli Yokulinganisa Ukuphepha kwepharamitha engu-300M Efana noma Edlule Ukunemba Kwamamodeli 23–90x Usayizi Wayo
Njengoba izinhlelo zokusebenza ezisebenza ngamandla e-LLM zingena ekukhiqizeni – futhi njengoba ama-agent e-AI enza imisebenzi ebaluleke kakhulu njengokuphequlula iwebhu, ukubhala nokwenza ikhodi, nokusebenzisana nezinsizakalo zangaphandle – ukulinganisela kokuphepha sekube ngenye yezingxenye ezibiza kakhulu zesitaki.
Onjiniyela abaningi abakhiphe isistimu yokukhiqiza ye-LLM bayayazi inkinga: udinga ukuhlola yonke imininingwane yomsebenzisi ngaphambi kokuthi ifinyelele imodeli, kanye nempendulo ngayinye yemodeli ngaphambi kokuthi ifinyelele kumsebenzisi. Lokho kusho ukuthi imodeli yakho ye-Guardrail isebenza kuso sonke isicelo esisodwa, ngaso sonke isikhathi sengxoxo. I-Guardrail latency compounds. Izindleko zihlanganisa. Futhi isizukulwane samanje samamodeli we-open-source guardrail — LlamaGuard4 (12B), WildGuard (7B), ShieldGemma (27B), NemoGuard (8B) – wonke angamamodeli anedekhoda kuphela anezigidigidi zamapharamitha, akhelwe ukuguquguquka kodwa hhayi isivinini.
I-Fastino Labs ikhiphe i-GLiGuard, imodeli yokuhlola ukuphepha yomthombo ovulekile yezigidi ezingu-300 eklanyelwe ukubhekana nale nkinga ethile. I-GLiGuard ihlola izilinganiso zokuphepha eziningi ngephasi eyodwa, futhi kuzo zonke izilinganiso zokuphepha eziyisishiyagalolunye, ukunemba kwayo kufana noma idlula amamodeli aphindwe izikhathi ezingama-23 kuya kwezingu-90 ubukhulu bayo kuyilapho igijima ngokushesha izikhathi ezifika kwezingu-16.
Ukuze uqonde ukuthi yini eyenza i-GLiGuard ihluke, kuyasiza ukuqonda ukuthi kungani amamodeli akhona e-Guardrail ehamba kancane. Amamodeli amaningi amakhulu we-Guardrail akhelwe phezu kwezakhiwo ze-decoder-only transformer, akhiqiza izinqumo zawo zokuphepha ngokuzenzakalelayo, ithokheni eyodwa ngesikhathi – ngendlela efanayo imodeli yolimi enkulu eyenza impendulo kumlayezo wengxoxo.
Lo mklamo wawunengqondo lapho izidingo zokuphepha ziwuketshezi. Amamodeli e-decoder angatolika izincazelo zemisebenzi yolimi lwemvelo futhi azivumelanise nezinqubomgomo zokuphepha ezintsha ngaphandle kokuqeqeshwa kabusha. Kodwa isizukulwane se-autoregressive silandelana ngokwemvelo, okwenza sihambe kancane futhi sibize ngekhompyutha.
Kunenkinga yokuhlanganisa phezu kwalokho. Amamodeli amaningi e-Guardrail adinga ukuhlola okokufaka kuzo zonke izici eziningi zokuphepha: hlobo luni lokulimala olukhona, noma ngabe ukwaziswa komsebenzisi uzama ukudlula ukuqeqeshwa kokuphepha, ukuthi impendulo yemodeli ngokwayo ayiphephile, njalo njalo. Ngenxa yokuthi amamodeli e-decoder akhiqiza okukhiphayo ngokulandelana, lokhu kuhlola kuvamise ukukhiqizwa ngokulandelana, futhi kuhlanganiswe i-latency njengoba kuhlolwa imibandela eyengeziwe.
Ngamanye amazwi, i-architecture eyenza amamodeli e-decoder aguquguquke futhi yisakhiwo esiwenza abe yithuluzi elingalungile lalokho okuyinkinga yokuhlukanisa.
Okwenziwa I-GLiGuard Empeleni
I-GLiGuard imodeli encane esekwe kusifaki khodi ehlela kabusha ukulinganiswa kokuphepha njengenkinga yokuhlukanisa umbhalo kunenkinga yokukhiqiza umbhalo. Amamodeli wesifaki khodi acubungula konke okokufaka ngesikhathi esisodwa futhi akhiphe ilebula yesigaba esisodwa sesethi yamalebula angaguquki, kuyilapho amamodeli edikhoda ekhiqiza okukhiphayo kwawo ithokheni eyodwa ngesikhathi, kwesokunxele kuye kwesokudla.
Ukuqonda okubalulekile kwezakhiwo ukuthi i-GLiGuard iyisingatha kanjani imisebenzi eminingi ngasikhathi sinye. Esikhundleni sokukhiqiza amathokheni, i-GLiGuard ibhala ngekhodi kokubili umbhalo wokufakwayo nezincazelo zomsebenzi (amalebula) ndawonye. Lawa abe esenikezwa imodeli, ethola amaphuzu kuwo wonke amalebula ngesikhathi esisodwa ekudluleleni phambili okukodwa bese ibuyisela ilebula enamaphuzu amaningi kakhulu ngomsebenzi ngamunye. Ngenxa yokuthi yonke imisebenzi namalebula ekhandidethi yawo ayingxenye yokokufaka ngokwayo, ukuhlola ubukhulu bokuphepha obengeziwe akungezi ukubambezeleka; kumane kusho ukufaka amalebula engeziwe kokokufaka.
I-GLiGuard iqhuba imisebenzi yokulinganisa emine ngesikhathi esisodwa ekudluleleni okukodwa:
- Ukuhlukaniswa kokuphepha (kuphephile / akuphephile) — isetshenziswa kukho kokubili ukwaziswa komsebenzisi ngaphambi kokukhiqiza kanye nezimpendulo zemodeli emva kwesizukulwane.
- Ukutholwa kwesu le-Jailbreak kuwo wonke amasu angu-11, afaka umjovo osheshayo, i-roleplay bypass, ukukhishwa kweziyalezo, nobunjiniyela bomphakathi. Uma kutholwa noma yiliphi isu le-jailbreak, ukwaziswa kumakwa ngokuzenzakalela njengokungaphephile.
- Ukutholwa kwesigaba sokulimaza kuzo zonke izigaba ezingu-14 – udlame, okuqukethwe kwezocansi, inkulumo enenzondo, ukuchayeka kwe-PII, ulwazi olungaqondile, ukuphepha kwezingane, ukwephulwa kwe-copyright, nokunye. Okokufaka okukodwa kungabangela izigaba eziningi ngesikhathi esisodwa.
- Ukutholwa kokwenqaba (ukuthobela / ukwenqaba), kulandelelwa ngokwehlukana ukuze kusizwe ukukala ukwenqaba (uma imodeli yenqaba izicelo eziphephile) futhi kutholwe ukuthobelana okungamanga (lapho imodeli ibonakala ithobela kodwa ingakwenzi lokho). Uma ukwenqaba kutholwa, impendulo imakwa ngokuzenzakalelayo njengephephile.
Idatha Yokuqeqesha Nokulungisa Kahle
I-GLiGuard yaqeqeshelwa ingxube yedatha yokuqeqeshwa echazwe ngumuntu kanye neyakhiwe ngokwenziwa. Ukuze uthole ukuphepha okusheshayo, ukuphepha kokuphendula, nokutholwa kokwenqaba, ithimba lasebenzisa i-WildGuardTrain, isethi yedatha yezibonelo ezingu-87,000 ezichazwe ngabantu. Ngesigaba sokulimala nokutholwa kwamasu e-jailbreak, amalebula amasampuli angaphephile akhiqizwa kusetshenziswa i-GPT-4.1.
Phakathi nokuqeqeshwa kwangaphambi kwesikhathi, imodeli yazabalaza ukuhlukanisa phakathi kwezigaba zokulimala okufanayo njengenkulumo enobuthi nodlame, ngakho ithimba lasebenzisa i-Pioneer ukukhiqiza idatha yokwenziwa yokwenziwa enamacala asemaphethelweni aqondise lokhu kuhlukaniswa okucolisekile.
Ngasohlangothini lwezokwakha, i-GLiGuard yaqeqeshwa ngokulungiswa kahle okugcwele kwendawo yokuhlola ye-GLiNER2-base-v1 izikhathi ezingu-20 kusetshenziswa i-AdamW optimizer. I-GLiNER2 iyisakhiwo sika-Fastino ngokwakhe sokuhlukaniswa kombhalo wemisebenzi eminingi — indawo yokuqala yemvelo yemodeli eklanyelwe ukuthola amasethi amalebula amaningi ngephasi eyodwa.
Imiphumela Yebhentshimakhi: Ukunemba Nesivinini
Ithimba labacwaningi lihlole i-GLiGuard kuwo wonke amabhentshimakhi okuphepha ayisishiyagalolunye amisiwe. Lawa mabhentshimakhi ahlanganisa kokubili ukuhlukaniswa okusheshayo nokuphendulayo, ukuhlola ukuthi imodeli ingakwazi yini ukuhlonza okuqukethwe okuyingozi, imelane nokuhlaselwa kwezitha, ihlukanise phakathi kwezinhlobo ezihlukene zokulimala, futhi igweme ukuhlaba umkhosi ngokweqile okuqukethwe okuphephile. Imiphumela isebenzisa i-macro-avareji ye-F1, imethrikhi evamile ebhalansisa ukunemba nokukhumbula.
Ngokunemba:
- I-GLiGuard ithola u-87.7 isilinganiso esingu-F1 ekuhlukaniseni ngokushesha, ngaphakathi kwamaphoyinti angu-1.7 wemodeli ehamba phambili (PolyGuard-Qwen ku-89.4).
- Ifinyelela isilinganiso sesibili esiphezulu se-F1 ekuhlukaniseni izimpendulo (82.7), ngemuva kwe-Qwen3Guard-8B kuphela (84.1).
- Idlula i-LlamaGuard4-12B, ShieldGemma-27B, ne-NemoGuard-8B naphezu kokuthi incane ngo-23–90×.
Ekuphumeni nasekubambeni isikhathi, imakwe ku-NVIDIA A100 GPU eyodwa:
- I-GLiGuard ifinyelela kufikela kokungu-16.2× kokuphuma okuphezulu (133 vs. 8.2 amasampula/s kusayizi weqoqo 4).
- I-GLiGuard ifinyelela ku-16.6× ukubambezeleka okuphansi: 26 ms vs. 426 ms ngobude bokulandelana okungu-64.
Lokhu akukona ukuthuthukiswa okuncane. Ku-26 ms ngesicelo ngasinye uma kuqhathaniswa no-426 ms, umehluko unengqondo kunoma yiluphi uhlelo lokusebenza lwesikhathi sangempela olubhekene nomsebenzisi, futhi umphumela ohlanganisayo engxoxweni enamajika amaningi wenza igebe likhulu nakakhulu ekusebenzeni.
Isichazi Esibonakalayo sikaMarktechpost
Okuthathwayo Okubalulekile
- I-GLiGuard iyimodeli yokulinganisa ukuphepha esekelwe kupharamitha engu-300M elawula imisebenzi emine – ukuhlukaniswa ngezigaba kokuphepha, ukutholwa kwekhefu lejele, ukuhlukaniswa kokulimala, nokutholwa kokwenqaba – ngokudlula okukodwa okuya phambili.
- Ngokungafani namamodeli e-guardrail e-decoder kuphela akhiqiza izinqumo ngokuzenzakalelayo, i-GLiGuard ihlela kabusha ukulinganisela kokuphepha njengenkinga yokuhlelwa kombhalo, isusa ibhodlela le-latency elandelanayo.
- Imakwe ku-NVIDIA A100 GPU eyodwa, i-GLiGuard ifinyelela kokufika kokungu-16.2× okuphezulu kanye ne-latency ephansi engu-16.6× (26 ms vs. 426 ms) uma kuqhathaniswa namamodeli amanje e-SOTA afana ne-ShieldGemma-27B.
- Kuwo wonke amabhentshimakhi okuphepha ayisishiyagalolunye, i-GLiGuard ithola u-87.7 isilinganiso esingu-F1 ekuhlukaniseni ngokushesha kanye nama-82.7 ekuhlukaniseni izimpendulo – i-LlamaGuard4-12B esebenza kahle kakhulu, i-ShieldGemma-27B, ne-NemoGuard-8B naphezu kokuthi incane ngo-23–90×.
- Izisindo zemodeli ziyatholakala ngaphansi kwe-Apache 2.0 ku-Hugging Face (
fastino/gliguard-LLMGuardrails-300M), okwenza isetshenziswe ku-GPU eyodwa ngaphandle kwengqalasizinda esindayo.
Hlola Iphepha, Izisindo zemodeli ku-HF, I-GitHub Repo futhi Imininingwane yobuchwepheshe. Futhi, zizwe ukhululekile ukusilandela Twitter futhi ungakhohlwa ukujoyina wethu 150k+ ML SubReddit futhi Bhalisela ku Iphephandaba lethu. Linda! ukutelegram? manje ungasijoyina kuthelegramu futhi.
Udinga ukusebenzisana nathi ekuthuthukiseni i-GitHub Repo yakho NOMA Ikhasi Lobuso Lokugona NOMA Ukukhishwa Komkhiqizo NOMA I-Webinar njll.? Xhumana nathi
