Dzokera kuNhau
ChengetedzoAI Understanding muchidimbu

BanglaVeilGuard inoedza magaba ekuchengetedza mukati memhando nhanhatu dzeBangla

Bepa idzva rinounza BanglaVeilGuard, bhenji uye isingaremi kukurumidza gadhi yakagadzirirwa kuyedza mamodheru emutauro weBangla pane akajairwa, Romanized, code-mixed, ruzha uye dialectal mafomu. Vanyori vanorondedzera kubudirira kwakadzika kurwiswa muongororo yavo, asi zvakare vanoona kuti murindi anoramba-kuramba kumwe kunhuhwirira…

7 min readRead the primary source
Source-page capture accompanying BanglaVeilGuard tests safety gaps across six forms of Bangla
Primary-source documentKwakanyorwa
Muparidzi
arxiv.org
Source link
arxiv.orghttps://arxiv.org/abs/2608.21880
Source type
Gwaro rekutanga - chiziviso chepamutemo, bepa, faira, kana peji rebato rekutanga ratinoverenga zvakananga.
ContextNzwisisa izvi mumasekonzi makumi matanhatu

Tanga pano

Matemu akakosha

Memory (Agent Memory)
Yakachengetwa mamiriro mumiriri weAI anoshandisa pamatanho kana masesheni kuvandudza kuenderera.
Kupatsanurwa
Basa iro modhi inogovera yekuisa kune imwe kana akawanda akatemerwa chikamu.
Kuongorora Seti
Dhata rakachengetwa rinoshandiswa kuyera mhando yemhando mushure mekudzidziswa.
Zviedze iwe pachakoAI Models Inotsanangurwa Mibvunzo

Chii chaitika

Bepa rakatumirwa kuarXiv musi waAug. 22 rinopa BanglaVeilGuard, Bangla-yekutanga kuchengetedza benchmark uye kuremerwa nekurinda kwemhando dzemitauro mikuru. Inoongorora mitauro mitanhatu: yakajairwa Bangla, Romanized Bangla, Banglish, Bangla-Chirungu kodhi-kusanganisa, ine ruzha Bangla uye dialectal Bangla.

Vanyori vanosuma BanglaVeilGuard sezvikamu zviviri zvakabatana: compact chengetedzo bhenji uye lightweight kukurumidza gadhi. Iyo bhenji ine 2,366 mhando-yakasefa yekukurudzira, ine yakabata-kunze yekuongorora yakakamurwa ye354 yekukurudzira. Izvo zvinokurudzira zvinotora zvikumbiro zvisina kuchengetedzeka, zvikumbiro zvakachengeteka uye zvikumbiro zvakachengetedzeka, zvichibvumira iyo sisitimu kuti iongororwe kwete chete kuti inovharira zvinokuvadza asi zvakare kuti inochengetedza kuwana kune zviri pamutemo mibvunzo. Kwarinobva kunozivisa basa sebepa remapeji masere rakagamuchirwa paMusangano wechina wepasi rose weComputing Advancements uye rakatumirwa seyakadhindwa yearXiv musi waNyamavhuvhu 22, 2026.

Iyo benchmark yakagadzirirwa kutenderedza musiyano mumanyorerwo eBangla. Mafomu ayo matanhatu akajairwa Bangla, Romanized Bangla, Banglish, kodhi-yakasanganiswa Bangla-Chirungu, ruzha Bangla uye dialectal Bangla. Iyo dhizaini inoratidza chirevo chepakati chebepa chekuti kuongororwa kwekuchengetedza kunogona kunge kusina kukwana kana ikatora script imwe yakajairwa kana imwe gungano rezviperengo. Kwakabva hakutsanangure kuvharwa kwenzvimbo yezvinyorwa zvemutauro, maitiro akashandiswa kutsanangura zvikamu, kana kuti zvirevo zvakasarudzwa sei uye nekusefa mhando kupfuura kutaura kuti seti yaive yakasefa.

Murindi anoshandisa isiri-inoparadza yakawanda-yekuona maitiro, yekukurumidza-njodzi classifier uye chikumbaridzo pre-chizvarwa gedhi. Mukutaura kwekuita, iyo nzira inovheneka kukurumidza kunouya pamberi pechizvarwa uye haichinji huremu hweiyo modhi inodzivirirwa. Bepa rinoti iyo nzira inogona kushandiswa kune heterogeneous tarisiro modhi. Kwakabva hakutauri kana gadhi aripo sekodhi, ingani compute kana latency yainowedzera, chii chikumbaridzo chakasarudzwa mukuedza kwega kwega, kana maitiro aanoita kana vashandisi vakabatanidza nemaune mafomu ekunyora akawanda.

Pakati pemhuri dzakanangwa-muenzanisi dzakadomwa muchidimbu, vanyori vanotaura kuti kuchengetedzwa kunomhanya kwakaderedza budiriro yekurwisa pasi pekutsidzira mhinduro kubva pakati pe93.8% kusvika 100.0% kusvika 6.3% yeClaude Opus 4.8, BanglaLLama uye TituLLM. YeTigerLLM-1B, bepa rinoshuma 78.2% kunyatsoita uye 8.8% kurwisa-kubudirira mwero neBanglaVeilGuard. Kuyeuka kusina kuchengeteka kwemurindi kwakanzi 88.5%, zvakanyanya pamusoro peyakaongororwa yekukurumidza-chete magadhi egadhi. Izvi zvakashumwa zvabuda pamapepa, kwete kudzokorora kwakazvimirira.

Vanyori vanozivisawo mutengo: iyo sisitimu yakawandisa-inoramba zvimwe zvinyoro zvinokurudzira, kunyanya izvo zvakanyorwa mumutauro weDialectal kana ruzha Bangla. Kuwana ikoko kwakakosha nekuti chengetedzo inogona kudzikisa zvinokuvadza zvinobuda uku ichivharisa kushandiswa zviri pamutemo. Kwakabva kunoisa izvi senzira yekubatsira kuchengetedza, asi chirevo hachiyere mwero wekuramba wenhema kana kuuparadza nemutauro, modhi, mhando yekukurumidza kana kuomarara.

Kwakabva mashoko: arxiv.org ↗

Nei zvichikosha

Iro basa rinogadzirisa kushaya simba kunoshanda mukuyedzwa kwekuchengetedza: modhi inobata yakajairwa-script Bangla inogona kusapindura zvakachengeteka kuchikumbiro chimwe chete kana ichidudzirwa, yakanyorwa zvisirizvo, yakasanganiswa kodhi kana kunyorwa murejista yedunhu. Mhedzisiro yakashumwa inoratidza kuti kuongorora kwe-cross-script kunogona kuburitsa njodzi dzakapotswa neChirungu-centric kana standard-script bvunzo.

Mitauro yakasiyana-siyana inyaya yekuchengeteka kana modhi dzichishandiswa nevanhu vasingarambe vachinyora nenzira yakamisikidzwa. Chikumbiro chinokuvadza chinogona kududzirwa mumabhii echiLatin, akasanganiswa neChirungu, achichinjwa nezviperengo zvisina kurongwa kana kuratidzwa murejista yedunhu. Kana hurongwa hwekuchengetedza huchingoziva maitiro epamusoro aripo mune yakajairwa-script kudzidziswa uye data yekuongorora, kuita kwayo kunooneka kunogona kusamiririra maitiro ayo mukushandiswa kwakajairwa kwemitauro yakawanda. BanglaVeilGuard inoita kuti dambudziko rekuongorora rive nyaya yakananga yechidzidzo.

Mupiro webepa saka mune chikamu methodological. Panzvimbo pekubata Bangla sechikamu chimwe chete chekuisa, inokurudzira kuyedza mafomu akati wandei mune imwe bhenji uye nekushandisa normalization pamberi pekuisa njodzi. Izvo zvinogona kubatsira vanogadzira modhi kuona kana mutemo wekuramba wakasimba pakuchinja mune script uye zviperengo. Iyo nzira zvakare ine huremu huremu mune tsananguro yepepa: inoshanda senge pre-chizvarwa gedhi uye haidi kugadzirisa zviyero zvemodhi yakachengetedzwa. Kana mibairo yakashumwa ikawanda, dhizaini iyoyo inogona kuve yakakosha kumasangano asingakwanisi kudzidzisazve kana kugadzirisa modhi yese yavanoshandisa.

Kuderedzwa kwakashumwa kwekubudirira kwekurwisa kwakanyanya mukati megadziriro yevanyori. Pepanhau rinoti mhuri nhatu dzine mazita emhando dzakasimuka kubva pa93.8% -100.0% kurwisa budiriro pasina murindi kuenda ku6.3% nayo, nepo TigerLLM-1B yakawana yakaderera yakashumwa kurwisa-kubudirira pamwero we78.2% chokwadi. Iyo sosi zvakare inoshuma 88.5% kusachengeteka kuyeuka. Nhamba idzodzo dzinoratidza kuti murindi anogona kubata akawanda asina kuchengetedzeka emhando dzakasiyana dzeBangla, asi ivo pachavo havaratidze kuti mamodheru ari pasi akachengeteka kana kuti dziviriro inoshingirira kurwiswa.

Iko kutengeserana kunokosha kune veruzhinji-yakatarisana masisitimu. Kunyanya-kuramba kunogona kurambidza vashandisi kuwana ruzivo rwakanaka, kunyanya kana mutauro wemutauro kana kuti une ruzha uchibatwa zvisirizvo seuri kufungira. Izvi zvinogona kukanganisa zvisingaite vanhu vane kunyora kwemazuva ese kusingaenderane nemabhenji akajairwa. Kwakabva hakugadzirise kugoverwa kwevanhu kwechikanganiso ichi, saka kushanda kwayo kunoramba kuri mubvunzo wakavhurika. Izvo, zvisinei, zvinopa humbowo hwekuti kunatsiridza kutariswa kwekuchengetedza uye kuchengetedza kubatsira matambudziko akabatana einjiniya pane kuparadzanisa zvinangwa.

Basa racho rakakoshawo kumubvunzo wakakura wekuti ongororo dzekuchengetedza dzinofanira kuratidza matauriro anoita vanhu. Sosi inotsigira mhedziso yakamanikana: bepa iri rinopa bhenji uye chengetedzo, uye vanyori vanoshuma zvakavandudzwa mhedzisiro pasi pekuongororwa kwavo kwakataurwa. Izvo hazviratidze kuti mamodheru ese eBangla ane kusagadzikana kwakafanana, kuti kuchinjika-manyoro ndiko kunonyanya kukonzeresa kukundikana kwekuchengetedza, kana kuti nzira yacho yaizoendesa kune mimwe mitauro pasina data nyowani uye kuyedzwa.

Interactive Mechanism

Interactive Mechanism: Iyo Inonyatsoshanda

Ongorora ari pasi tekinoroji kuseri kwekusimudzira uku uchipindirana.

Document Size:128K tokens
Needle Placement Depth (Location in document):50% into text
Attention Context Buffer Map:
Target Fact (50%)
Equivalent Pages~320Standard book pages
Retrieval Accuracy99.9%Needle recall score
RAM / KV Cache5.1 GBMemory overhead
Prompt CachingActive~80% discount on reuse
Core takeaway: Million-token context windows allow querying whole codebases or legal archives in one prompt. However, KV cache memory scales with context length, making prompt caching crucial for real-time production.
Interactive Concept Check+10 Points
AI Models Explained Quiz

Which component of an AI application is the machine-learning model itself?

Zvekutarisa zvinotevera

Mubvunzo wakakosha ndewekuti izvo zvakaburitswa zvinobata kunze kwekumisikidzwa kwepepa. Iyo sosi haipe ruzivo nezve iyo yakazara yaivavarirwa-modhi yekumisikidza, kuvaka kurwisa, yekutanga kuita, maitiro ekuisa zvibodzwa kana chaiyo-yepasirese deployment, uye inotaridza kuramba-kuramba pane benign dialectal uye ruzha rumbidzo semuganho usina kugadziriswa.

Kumwe kuongorora kunofanirwa kutanga nebhenji pachayo. Kunobva kunopa huwandu hwese hwekurudziro uye kupatsanurwa-kwakabatwa, asi kwete kugoverwa kwemienzaniso isina kuchengetedzeka, yakachengeteka uye ine hutsinye mumitauro mitanhatu. Izvo hazvitauri zvakare kuti mangani ekurudziro ari ega rega rega, kurwiswa kwakagadzirwa sei, kana kuti iyo yekuongorora yakaiswa yakazvimiririra yedata regadhi. Iwo mameseji anogona kukanganisa kuti zvakashumwa kurwisa-kubudirira uye nhamba dzekurangarira dzinogona kududzirwa sei.

Iyo evaluation protocol ndeimwe yakakosha isingazivikanwe. Zvigumisiro zvinoshandisa deterministic response scoring, asi kwakabva hakutsananguri scoring rubriki, tsanangura kana mhinduro dzakatongwa otomatiki kana nevanhu, kana kuratidza kuti kurambwa kwemuganhu kwaiitwa sei. Izvo zvakare hazvitsananguri akaongororwa ekukurumidza-chete magadhi ekutanga. Kuzviyedza kwakazvimirira kwaizobatsira, kunyanya nesampling isina kurongeka, zvirevo, dudziro isingaonekwe, mataurirwo asina kumiririrwa mudata rebudiriro uye zvirevo zveanopikisa zvakagadzirirwa mushure mekuburitswa kwemurindi.

Mhedzisiro yemuenzaniso inoda kupatsanurwa kubva kune zvakajairika nezvekuchengetedza kwemuenzaniso. Iwo abstract mazita Claude Opus 4.8, BanglaLLama uye TituLLM, uye inopa yakaparadzana mhedzisiro yeTigerLLM-1B, asi haipe mavhezheni emhando, mamiriro ekuwana, sisitimu inosimudzira, decoding marongero kupfuura deterministic scoring, kana kugovera chaiko kwemabasa. Iyo sosi zvakare haitauri latency, ndangariro kushandiswa, mutengo wekushandisa kana kuwanikwa. Izvo zvakasiiwa zvinosiya kusava nechokwadi nezvekuti nzira yacho inogona kubatanidzwa zviri nyore sei mumasisitimu ekugadzira.

Nyaya yehunyanzvi yekukurumidza kuramba. Vanyori vanonyatso ratidza benign dialectal uye mheremhere sekushaya simba, asi kwazvinoita haaverenge chikanganiso kana kuratidza kana kuchinja kwechikumbaridzo kungavandudza chiyero. Ongororo dzenguva yemberi dzinofanira kuzivisa kuyeuka nezvekuchengetedzwa pamwe chete nekugamuchirwa kwechimbichimbi nechimiro chemutauro, uye kunofanirwa kuyedza kana kuita kuti kugarisa pachakwo kuchidzima misiyano yemitauro ine musoro kana kushandura chinangwa chekukurumidza.

Pakupedzisira, chimiro chepepa uye chiyero chinofanira kuramba chakajeka. Ndiyo arXiv shanduro yakatumirwa muna Aug. 22 uye peji inoti yakagamuchirwa paICCA 2026; iyo sosi haipe yakazvimirira kudzokorora kana humbowo hwekutumirwa. Nhamba dzakataurwa dzinobatwa zvakanaka semhedzisiro kubva kune imwe ongororo yekuongorora. Chauchatarisa chinotevera kodhi kana kuburitswa kwedata, kubereka kwakazvimirira, kuyedzwa kurwiswa kunoshanduka uye kuyerwa kwakafara kwekubatsira mumhando dzakasiyana dzeBangla.

Related guides & Quizzes

AI Models InotsanangurwaTsika dzeAIKudzidziswa kweAIPrompt EngineeringEdza zvaunoziva - edza yemahara AI quizTarisa kumusoro izwi reAI mune yedu glossaryTevedza iyo AI regulation tracker
Wakawana izvi zvinobatsira?