1. Gabatarwa & Bayanin Matsala
Mafi yawan tsarin horar da ƙananan model na harshe masu inganci (dalibai) ya ƙunshi jagora daga manyan model masu iyawa (malamai). Duk da haka, wannan hanya ta ci karo da babban cikas: rashin daidaituwar ƙamus. Lokacin da malami da model ɗin dalibi suka yi amfani da masu rarraba alamomi daban-daban—wani yanayi na gama gari lokacin amfani da model ɗin buɗaɗɗen tushe ko na musamman—jerin alamominsu da rarraba yuwuwar fitarwa sun bambanta, suna lalata ingantaccen canja ilimi. Kamar yadda aka nuna a cikin takarda, babban model kamar Qwen2.5-Math na iya raba kusan 6.32% kawai na ƙamus ɗinsa tare da dalibi kamar TinyLlama, yana haifar da babban cikas ga amfani da mafi kyawun model ɗin da ake da su a matsayin malamai.
2. Tsarin VocAgnoLM
Model na Harshe Jagorar Malami maras ƙamus (VocAgnoLM) ya ba da shawarar mafita mai kafa biyu don cike wannan gibi, yana ba da damar tace ilimi maras ƙamus.
2.1 Babban Fahimta & Tsarin Ma'ana
Babban Fahimta: Babban cikas ba tsarin model ba ne, amma rashin daidaitawar wakilci. Ba za ku iya kwatanta apples (alamomin Qwen) da lemu (alamomin TinyLlama) kai tsaye ba. Hazakar VocAgnoLM ta ta'allaka ne a sake tsara matsalar daga "daidaita abubuwan da aka fitar" zuwa "daidaita sararin ma'ana da alamun koyo." Yana raba ilimin malami daga tsarin rarraba alamominsa na musamman.
Tsarin Ma'ana: Tsarin yana da tsari mai kyau: 1) Don rubutu da aka shigar, samar da jerin alamomi don model ɗin dalibi da malami. 2) Yi amfani da Daidaitawar Ƙamus a Matakin Alama don ƙirƙirar taswira tsakanin jerin da ba su dace ba. 3) Yi amfani da wannan taswirar don amfani da Asarar Jagorar Malami, ta yin amfani da asarar cikin malami a matsayin alamar horo ga dalibi, ta ƙetare daidaitawar yuwuwar alama kai tsaye.
2.2 Daidaitawar Ƙamus a Matakin Alama
Wannan ɓangaren yana magance matsalar rashin daidaitawar jerin. Yana kafa taswira ɗaya-zuwa-da-yawa daga kowane alamar dalibi zuwa jerin alamomin malami da suka dace. Misali, alamar dalibi "Pro" na iya yin taswira zuwa alamomin malami "Prob" da "ability". Wannan a zahiri yayi kama da dabarun daidaitawa a cikin fassarar inji (kamar waɗanda ake amfani da su a cikin MT na ƙididdiga ko farkon model na jijiyoyi) amma ana amfani da su a matakin ƙaramin kalma a cikin tsare-tsaren rarraba alamomi daban-daban. Manufar ita ce ƙirƙirar gada wanda zai ba da damar gudanar da bayanai duk da rabuwar ƙamus.
2.3 Asarar Jagorar Malami
Maimakon tilasta wa dalibi ya kwaikwayi rarraba yuwuwar alama ta gaba na malami—wanda ba zai yiwu ba tare da ƙamus daban-daban—VocAgnoLM yana amfani da asarar ƙirar harshe na malami da kansa a matsayin jagora. An horar da dalibi don rage haɗakar manufa: daidaitaccen asarar ƙirar harshensa da kuma asarar da ke ƙarfafa wakilcinsa na ciki ko tsinkaya don haifar da ƙananan ƙimar asara ga model ɗin malami akan jerin da aka daidaita. Wannan wani nau'i ne mai zurfi, amma mai ƙarfi, na jagora.
3. Ƙarfafawa & Kurakurai Masu Muhimmanci
Ƙarfafawa:
- Yana Buɗe Bambancin Model: Wannan shine fasalin mai kashewa. Yana karya kullewar mai siyarwa/tsarin muhalli, yana ba da damar ƙungiyoyi su yi amfani da mafi kyawun model ɗin da ake da su (misali, Qwen na musamman na lissafi) don koyar da kowane dalibi, ba tare da la'akari da asalinsa ba (misali, TinyLlama).
- Mai Aiki & Mai Sauƙi: Ba ya buƙatar sake horar da mai rarraba alamomin malami ko ɓangaren haɗakar dalibi, yana guje wa babban nauyin injiniya.
- Sakamako Mai Ƙarfi na Gwaji: Haɓaka aikin kashi 46% akan horon farko maras hankali tare da mummunan rashin daidaituwar ƙamus ba ƙaramin abu bane. Yana nuna hanyar tana aiki a aikace.
Kurakurai Masu Muhimmanci & Tambayoyin Buɗe:
- Daidaitawar Heuristic Akwatin Baƙi ne: Takardar ta yi watsi da ainihin algorithm don "Daidaitawar Ƙamus a Matakin Alama." Shin tsarin shirye-shirye ne mai ƙarfi? Model ɗin da aka koya? Ƙarfin ƙarfi da farashin lissafi na wannan matakin daidaitawa sune mahimman abubuwan da ba a sani ba. Mummunan daidaitawa zai iya yada hayaniya maimakon ilimi.
- Asarar Alamar Fine-Grained: Yin amfani da asarar scalar na malami yana sadaukar da arziƙin, babban siginar girma na cikakken rarraba fitarwa. Yana kama da koyo daga maki na ƙarshe maimakon cikakken bayani akan kowane amsa. Wannan na iya iyakance amincin canja ilimi don iyawar harshe mai zurfi.
- Scalability zuwa Mummunan Rashin Daidaituwa: Rashin daidaituwar da aka gwada (6% haɗuwa) yana da tsanani, amma game da kusan sifili haɗuwa? Iyakokin ka'idar wannan hanyar ba a gwada su ba.
4. Sakamakon Gwaji & Bincike
4.1 Tsari & Ma'aunin Aiki
Binciken yana amfani da model ɗin dalibi mai sigogi 1B (TinyLlama) da model ɗin malami 7B daban-daban (Llemma, Mistral, DeepSeek-Math, Qwen2.5-Math) tare da girman ƙamus daga 32K zuwa 150K. Babban ma'auni shine aiki akan jerin kimanta lissafi, kwatanta VocAgnoLM da ma'auni na ci gaba da horon farko ba tare da jagorar malami ba.
4.2 Babban Binciken & Fassarar Ginshiƙi
Babban sakamakon ana ganinsa a cikin Hoto na 1 na takarda. Yana nuna mahimman abubuwa guda biyu:
- Matsalar Rashin Daidaituwar Ƙamus: X-axis yana nuna model ɗin malami tare da haɓaka aiki (daga Llemma zuwa Qwen2.5-Math). Ginshiƙin yana nuna haɗuwarsu na ƙamus tare da TinyLlama. Akwai bayyanannen dangantakar sabanin: mafi kyawun malami (Qwen) yana da mafi ƙarancin haɗuwa (~6%). Wannan yana nuna matsalar da VocAgnoLM ke nufin warwarewa.
- Ingancin VocAgnoLM: Rubutun ya bayyana cewa tare da Qwen2.5-Math a matsayin malami, VocAgnoLM ya sami haɓaka aikin kashi 46% akan ma'auni. Wannan ya tabbatar da cewa tsarin ya yi nasarar amfani da malami mai ƙarfi duk da ƙaramin haɗuwar ƙamus. Takardar ta kuma lura da fa'idodi masu daidaito daga malamai masu ƙarfi, yana tabbatar da ainihin jigon.
Babban Sakamakon Gwaji
Haɓaka Aikin Kashi 46% da VocAgnoLM ya samu ta amfani da Qwen2.5-Math (haɗuwar ƙamus 6.32%) a matsayin malami ga TinyLlama, idan aka kwatanta da daidaitaccen ci gaba da horon farko.
5. Hanyoyin Aiki & Tasirin Dabarun
Ga masu aiki da shugabanni a cikin AI:
- Dabarun Nan Take: Idan kuna gina model na musamman (misali, don kuɗi, doka, ilimin halittu), daina iyakance binciken malamanku ga model ɗin da ke da masu rarraba alamomi masu dacewa. A yi kimanta mafi kyawun model ɗin a cikin yankinku, ba tare da la'akari da mai rarraba alamominsu ba. VocAgnoLM yana ba da hanya mai yuwuwa don amfani da su.
- Sayayyar Dabarun: Wannan binciken yana rage haɗarin "kullewar mai rarraba alamomi." Lokacin zaɓar tushen model don ƙungiyarku, daidaitawar ƙamus ya zama ƙaramin takura, yana 'yantar da ku don zaɓa bisa tsarin gine-gine, lasisi, da aiki kawai.
- Zuba Jari na Bincike: Bangaren daidaitawa shine maɓalli. Zuba jari a cikin ingantattun, ingantattun, kuma mai yuwuwar hanyoyin daidaitawa da za a iya koyawa zai zama mabuɗin masana'antar wannan hanyar. Yi la'akari da shi a matsayin iyaka na gaba a cikin haɗin kai na model.
- Takaici: Wannan ba harsashi na azurfa bane. Don ayyukan da ke buƙatar ingantaccen samarwa ko kwaikwayon salo, asarar daidaitawar rarraba mai zurfi na iya zama babban nakasa. Gwada shi don ayyukan da ke da ilimi mai yawa (kamar lissafi, tunani) da farko.
6. Zurfin Fasaha
6.1 Tsarin Lissafi
Duk da yake ba a bayyana cikakken aikin asara ba a cikin abin da aka fitar, ana iya tsara ainihin ra'ayi. Bari $\mathcal{V}_s$ da $\mathcal{V}_t$ su zama ƙamus ɗin dalibi da malami. Don jerin shigarwa $x$, dalibi yana samar da jerin alama $\mathbf{s} = [s_1, ..., s_n]$ kuma malami yana samar da $\mathbf{t} = [t_1, ..., t_m]$, tare da $n \neq m$ gabaɗaya.
Aikin Daidaitawar Ƙamus a Matakin Alama $\mathcal{A}$ yana yin taswira kowane alamar dalibi $s_i$ zuwa jerin alamomin malami masu ci gaba: $\mathcal{A}(s_i) = \mathbf{t}_{[j:k]}$.
Asarar Jagorar Malami $\mathcal{L}_{guide}$ mai yiwuwa ya ƙunshi ciyar da wakilci ko tsinkaya da aka samu daga dalibi (wanda aka daidaita ta hanyar $\mathcal{A}$) cikin ci gaba na malami kuma a lissafta asarar ƙirar harshe na malami akan shi. Manufar horon dalibi ta zama:
$$\mathcal{L}_{total} = \mathcal{L}_{LM}(\theta_s; x) + \lambda \cdot \mathcal{L}_{guide}(\theta_s, \theta_t; x, \mathcal{A})$$
inda $\theta_s$ da $\theta_t$ su ne sigogi na dalibi da malami, $\mathcal{L}_{LM}$ shine daidaitaccen asarar ƙirar harshe na dalibi, kuma $\lambda$ shine hyperparameter mai auna nauyi. Mahimmin abu shine cewa $\mathcal{L}_{guide}$ yana aiki akan jerin da aka daidaita, yana kewaye da rashin daidaituwar ƙamus kai tsaye.
6.2 Tsarin Bincike: Nazarin Lamari
Yanayi: Kamfani yana son ƙirƙirar ƙaramin, ingantaccen LLM don nazarin takaddun shari'a. Mafi kyawun malami na musamman da ake da shi shine `LexLaw-70B`, wanda ke amfani da mai rarraba alamomi na al'ada da aka horar da shi akan tarin rubutun shari'a. Dalibin da aka yi niyya shine model ɗin `Llama-3-8B`.
Aikace-aikacen Tsarin:
- Binciken Matsala: Bincika haɗuwar ƙamus. Yana da yuwuwar ƙasa da 20%. Tace ilimi kai tsaye ba zai yiwu ba.
- Lokacin Daidaitawa: Gudanar da samfurin rubutun shari'a ta hanyar model ɗin biyu. Yi amfani da ɓangaren daidaitawa na VocAgnoLM (misali, algorithm mafi ƙarancin gyara nisa akan rufaffiyar haɗin gwiwar byte) don gina taswira $\mathcal{A}$ tsakanin alamomin Llama-3 da jerin alamomin LexLaw don sharuɗɗan shari'a na gama gari (misali, "force majeure").
- Lokacin Horarwa: Horar da dalibin Llama-3 akan tarin rubutun shari'a. Don kowane rukuni, lissafta daidaitaccen asararsa. A layi daya, don kowane jerin, yi amfani da $\mathcal{A}$ don gina "kallon malami" na jerin da dalibi ya tsinkaya, mika shi zuwa ga malami LexLaw daskararre, kuma a lissafta asararsa. Mayar da haɗakar asarar don sabunta sigogin dalibi kawai.
- Kimantawa: Kula da aiki akan ma'auni na QA na shari'a akan ma'auni na dalibi da aka horar ba tare da jagorar LexLaw ba. Sakamakon da ake tsammani shine ingantaccen tunanin shari'a ba tare da canza mai rarraba alamomin dalibi ba.
7. Aikace-aikacen Gaba & Hanyoyin Bincike
- Canja Hanyoyin Cross-Modal & Cross-Lingual: Ainihin ka'idar daidaita sararin wakilci daban-daban tana da mahimmanci. Aikin gaba zai iya ƙaddamar da wannan don amfani da malami na harshe na gani (kamar GPT-4V) don jagorantar dalibi na rubutu kawai ta hanyar daidaita nau'ikan taken-hoto, ko amfani da malami mai albarkatu masu yawa don jagorantar dalibin harshe mai ƙarancin albarkatu.
- Daidaitawar Mai Ƙarfi & Da aka Koya: Matsawa daga daidaitawar heuristic zuwa ƙaramin, model ɗin daidaitawa da za a iya horarwa wanda ke koyon mafi kyawun taswira yayin horo zai iya inganta ƙarfi da inganci.
- Bututun Model na Masana'antu: Wannan yana ba da damar ƙirƙirar "kasuwannin malamai" inda ƙungiyoyi za su iya ba da daskararrun, model ɗin malami na musamman a matsayin sabis. Masu amfani na ƙasa za su iya tace waɗannan zuwa tsarin gine-ginensu na zaɓi, suna kare IP (malamai suna daskararre) kuma suna tabbatar da dacewa.
- Koyo na Tarayya tare da Abokan Ciniki daban-daban: A cikin yanayi na tarayya, abokan ciniki na iya amfani da tushen model daban-daban. VocAgnoLM zai iya samar da hanya don tattara ilimi daga waɗannan model ɗin daban-daban zuwa cikin model ɗin duniya ba tare da buƙatar daidaitawa ba.
8. Nassoshi
- Shin, H., Ji, L., Liu, X., & Gong, Y. (2025). Shawo Rashin Daidaituwar Ƙamus: Model na Harshe Jagorar Malami maras ƙamus. arXiv preprint arXiv:2503.19123.
- Zhang, P., et al. (2024). TinyLlama: Ƙaramin Model na Harshe na Buɗaɗɗen Tushe. Ma'ajiyar GitHub.
- Yang, A., et al. (2024). Qwen2.5-Math: Jerin Manyan Model na Harshe don Magance Matsalolin Lissafi. Rahoton Fasaha.
- Hinton, G., Vinyals, O., & Dean, J. (2015). Tace Ilimi a Cikin Hanyar Sadarwar Jijiyoyi. arXiv preprint arXiv:1503.02531. (Aikin farko akan tace ilimi).
- Zhu, J.-Y., Park, T., Isola, P., & Efros, A. A. (2017). Fassarar Hotuna-zuwa-Hoto mara Haɗin gwiwa ta amfani da Cibiyoyin Sadarwar Motsa jiki masu Daidaituwa. Proceedings of the IEEE International Conference on Computer Vision (ICCV). (Aiki mai tasiri akan daidaita rarraba a cikin yankuna daban-daban, kwatankwacin ƙalubalen daidaitawa a nan).
- Google AI. (2023). Gemma: Buɗaɗɗen Model Dangane da Bincike da Fasahar Google. https://ai.google.dev/gemma.
- Meta AI. (2024). Llama 3 Model Card. https://llama.meta.com/llama3/.