Chii chaitika
Chinyorwa chitsva chearXiv chinoongorora kuti post-training inokanganisa sei kunzwisisa mutauro weBangla mumhuri nhatu dzemhando dzemitauro mikuru. Vanyori vanofananidza yakazara-chaiyo uye matatu quantized mafomati mukati meshanu Bangla echisikigo-mutauro-kunzwisisa mabhenji.
Pepa racho, rakatumirwa kuarXiv muna Aug. 25, rinoongorora post-training , nzira inoshandiswa kuderedza chiyeuchidzo chinodiwa nemitauro mikuru yemitauro uye kukurumidza kufungidzira. Vanyori vanogadzira mubvunzo wakatenderedza Bangla, mutauro wavanotsanangura se morphologically yakaoma uye yakaderera-sosi, vachipokana kuti kwakawanda kusati kwanzwisisa kwehuwandu kunobva kumabhenji echiRungu. Mupiro wakataurwa wechidzidzo uyu kuenzanisa kwakadzorwa kwehuwandu hwemafomu ekunzwisisa Bangla-mutauro wechisikigo. Mune mamwe mazwi, chidzidzo chacho chakagadzirirwa kuenzanisa kufanana-kwe-kufanana nemuenzaniso wekuongorora uku uchishandura nhamba inomiririra inoshandiswa panguva yekufungidzira.
Ongororo iyi inosanganisira mhuri nhatu dzemhando: Qwen-2.5-7B, LLaMA-3.1-8B uye GPT-OSS-20B. Imwe neimwe inoongororwa mukunyatso kuzere uye mumatatu ekumisikidzwa akaonekwa mune abstract seGPTQ-Int8, GPTQ-Q8 uye GGUF-W8A16. Iwo maedzo anoshandisa zero-shot ongororo kuburikidza nelm-evaluation-harness framework uye span mabenchmark mashanu: Bangla MMLU, CommonsenseQA-BN, OpenBookQA-BN, PIQA-BN uye BoolQ-BN. Tsime rinoti bepa rine mapeji masere, tafura imwe uye apendix imwe, asi rekodhi rakapihwa harisanganisire tafura yemhedzisiro. Iyi setup inobvumira kuenzanisa kuongorora modhi-yemhuri maitiro uye mutsauko pakati pemafomati ane mazita achipesana neamwechete akataurwa bhenji.
Chinonyanya kutaurwa chinowanikwa ndechekuti mhuri dzemuenzaniso dzinopindura zvakasiyana kune . GPT-OSS yakarasika kusvika 57.35% kunyatsoita pakufunga-anorema mabasa pasi peGGUF-W8A16. Iyo abstract inoti Qwen uye LLaMA yakaramba yakatsiga pasi peGPTQ, uye kuti mavhezheni ehuwandu akadarika mhedzisiro yakazara mune mashoma kesi. BoolQ-BN, inotsanangurwa sebasa rekunzwisisa, yakaramba yakagadzikana kumhuri dzese dzemhando nhatu uye mafomati. Mhedzisiro yacho saka patani yebasa- uye fomati-chaiyo shanduko pane imwe nzira yekuchinja pachidzidzo chose.
Izvi zvirevo zvakaitwa nepreprint; tsime haripe ruzivo rwakakwana pano kuti uone kurongeka kwekutanga, shanduko yebasa-ne-basa chaiyo kana kuti kuderera kukuru kunomiririra hama kana kurasikirwa kwechikamu. Izvi zvinoganhura kuti mugumisiro wenhamba unogona kududzirwa sei kubva pane zvakapihwa chete.
Nei zvichikosha
Mhedzisiro yacho inoratidza kuti kudzikisa iyo AI modhi yekurangarira tsoka haiburitse yunifomu mumitauro, zvivakwa kana mabasa. Kumasangano anoshandisa mamodheru emitauro pane zvakamanikidzika, modhi uye sarudzo dzinogona kukosha sezita rezita.
Nyaya inoshanda ndeyekuti inowanzo tariswa senge sarudzo inoshanda: shandisa mashoma mabhiti kudzikisa ndangariro kushandiswa uye zvinogona kuwedzera inference kumhanya. Iri bepa rakashumwa mhedzisiro inoratidza kuti, yeBangla, mutengo wemhando inogona kuenderana nekudyidzana pakati peiyo modhi yekuvaka, quantization nzira uye basa. Sarudzo yekuendesa inoita seyakagamuchirwa pane imwe bhenji kana mhuri yemuenzaniso inogona kuburitsa mhedzisiro yakasiyana pane imwe. Sarudzo yacho inosungirwa kune basa ririkushandirwa, kwete kungochengeterwa kana kumhanyisa chinangwa.
Zvakawanikwa zvinonyanya kukosha pakufunga-zvinorema zvikumbiro, nekuti kushatisa kukuru kwakashumwa kunoitika muchikamu ichocho chekuongorora kwete zvakafanana pamabasa ese. Panguva imwecheteyo, kugadzikana kweBoolQ-BN kunoratidza kuti nei mhedziso yakafara nezve "" ingave inotsausa. Iyo sosi inopa yakasanganiswa patani: mamwe emhando-fomati musanganiswa anodzikisira, mamwe anoramba akagadzikana uye mamwe anonzi anovandudza zvishoma. Izvo zvinoita kuti bhenji-chaiyo yekuyedza ikoshe pane kuvimba nehupamhi hwega. Mukutaura, mubairo wakanaka pane imwe chidimbu chekuongorora haungagone kusimbisa mamiriro akafanana kune imwe nzvimbo.
Mupiro wakakura ndewe kuyerwa. Vashandisi veBangla uye vanogadzira vanogona kusashandirwa zvakanaka nekufungidzira kuti mhedzisiro kubva mumutauro weChirungu-yekuongorora inotamisa zvakananga. Iro pepa hariratidzi kuti quantized modhi haina kukodzera Bangla deployment; mhedziso yaro yakamanikana uye inonyanya kubatsira: inogona kushanda, asi yekuvaka uye quantization fomati inoda kusarudzwa nemutauro unotariswa uye basa mupfungwa. Iyo framendi inochengeta chirevo chebepa chakanangana nekuongorora uye kusarudzwa kwete pamutongo wegumbeze nezve kuderedzwa chaiko.
Hapana humbowo mune kwakapihwa hunoratidza mhedzisiro kune vashandisi vekugadzira, kuchengetedza zvabuda, mhando yeshanduro, masisitimu ekutaura kana mamwe maapplication kunze kwezvishanu zvakanyorwa mabhenji. Iyo mibvunzo inoramba iri kunze kweiyo mabhenji akapihwa anogona kupindura.
Interactive Mechanism: Iyo Inonyatsoshanda
Ongorora ari pasi tekinoroji kuseri kwekusimudzira uku uchipindirana.
Which component of an AI application is the machine-learning model itself?
Zvekutarisa zvinotevera
Iyo yakakosha yekutevera ndeyekuti iyo yakashumwa pateni inobata kune mamwe mamodheru, Bangla mabasa uye marongero ekutumira. Sosi yacho inozivisa basa searXiv vhezheni 1 kutumira uye haitsigire ongororo yevezera, kudzokorora kwakazvimiririra kana chaiyo-nyika mushandisi maitiro.
Mubvunzo wekutanga ndeye reproducibility. Vanyori vanotsanangura basa seyekutanga inodzorwa kuenzanisa mafomati paBangla natural-language nzwisiso, asi rakapihwa arXiv rekodhi haritauri kana kodhi, modhi mafaera, calibration data kana yakazara ongororo yakabuda. Kudzokororwa kwakazvimirira kwaizobatsira kuona kana iyo yakashumwa 57.35% kurasikirwa kwakanyanya kwakasimba kune kuita sarudzo uye magadzirirwo ekuongorora. Tsananguro iripo saka inotsigira chikumbiro chezvigadzirwa uye kudzokorora, pane mhedziso yekuti mhedzisiro yacho inogona kudhindwa here.
Mubvunzo wechipiri ndewe scope. Iyo abstract haipe huwandu hwemibvunzo mubenchmark yega yega, nguva dzekusavimbika, mutsauko pane zvirevo kana zvibodzwa zvebasa. Izvo zvakare hazvitsananguri maitiro ekugadzirisa kuseri kweyega yega quantized modhi, iyo inogona kukanganisa kuenzanisa. Izvo zvakasiiwa zvinoreva kuti iyo yakanyanya kutaurwa haifanirwe kuve yakajairwa kune ese maBangla ekushandisa kesi kana kubatwa sechirango chepasi rose cheGGUF-W8A16. Mamiriro ekushaikwa akakosha nekuti huwandu hwepamusoro pakuenzanisa kwakataurwa hunogona kusatsanangudza mhedzisiro.
Rimwe basa rinofanirwa kuyedza mamwe mamodheru mhuri, zvirongwa uye Bangla mabhenji, kusanganisira anoshanda akawandisa senge chizvarwa, rairo rinotevera uye refu-fomu mhinduro kana vaongorori vakasarudza kuzvidzidza. Inofanirawo kuongorora kana kuwana kana kurasikirwa kuchienderera mberi kwemhando yemhando uye hardware. Kuwedzerwa kwakadaro kwaizojekesa kana iyo iripo benchmark pateni mepu pane yakakura inoshandisa vanogadzira vane hanya nazvo.
Kwakabva kunozivisa iyi searXiv vhezheni 1, yakatumirwa Aug. 25, uye hairatidze wongororo yevezera rako kana kusimbiswa kwekunze. Kusvika iwo macheki aripo, bepa rinonyanya kuverengwa seyakatarisana nekuongorora iyo inosimudza kunetseka kwekutumira, kwete seyakajeka chinzvimbo cheanokwana Bangla-anokwanisa modhi. Chimiro ichocho chinofanira kuramba chiri chikamu chekuti mhedzisiro inodudzirwa sei nepo humbowo huchikura.