Gabatarwa

Ka yi tunanin yin amfani da ƙarfin samfuran ci-gaban harshe daidai akan kwamfutarka ko na'urar tafi da gidanka ba tare da dogaro da sabis na gajimare ko sabar masu ƙarfi ba. Sauti mai ban mamaki, ko ba haka ba? To, waɗannan ƙananan ƙirar harshe sun sa wannan mafarki ya zama gaskiya. A cikin NLP, mun lura da zuwan manyan nau'ikan harshe waɗanda ke haɗawa da ƙirƙirar rubutu kamar ɗan adam. Duk da yake sakamakon sau da yawa yana da ban mamaki, buƙatun ƙididdiga suna da girma daidai da girma. Sakamakon haka, yana da wahala a tafiyar da su a wajen cibiyar sarrafa su. Amma wannan yana canzawa da sauri! Labari mai dadi shine cewa masu bincike da injiniyoyi sun zube zukatansu don samar da ƙananan LLMs waɗanda suka isa suyi aiki akan na'urorin ku na gida kuma suna da isasshen iko don amfani da su ga kowane aiki mai amfani.

A cikin wannan labarin, za mu bincika mafi ƙanƙanta kuma mafi girman nau'ikan yare waɗanda zaku iya gudanar da su cikin gida daga jin daɗin na'urar ku. Waɗannan ƙaƙƙarfan abubuwan al'ajabi suna daidaita daidaitaccen daidaito tsakanin aiki da ingantaccen albarkatu, buɗe duniyar yuwuwar masu haɓakawa, masu bincike, da masu sha'awa iri ɗaya.

Menene Fa'idodin Ƙananan LLMs?

Anan akwai wasu mahimman fa'idodin amfani da ƙananan LLMs (Manyan Harshe Model) idan aka kwatanta da manyan takwarorinsu:

1. Ƙananan Bukatun Hardware: Ƙananan LLMs suna da ƙananan sigogi kuma suna buƙatar ƙarancin ikon lissafi, yana sa su dace don aiki akan na'urori masu iyakacin kayan aiki, kamar kwamfyutoci, wayoyin hannu, da tsarin da aka saka. Wannan yana ba su damar samun dama da haɓaka dimokraɗiyya ta amfani da LLMs don faɗuwar kewayon masu amfani da aikace-aikace.
2. Fitowar Fitowa: Tare da ƙananan sigogi da ƙananan girman samfurin, ƙananan LLMs na iya yin ƙima da sauri, wanda ke nufin lokutan amsawa da sauri da ƙananan latency. Wannan yana da mahimmanci musamman ga aikace-aikacen lokaci-lokaci kamar tattaunawa AI, inda amsawa ke da mahimmanci.
3. Ƙananan Amfanin Makamashi: Ƙananan samfura suna buƙatar ƙarancin makamashi don gudu, yana sa su zama masu amfani da makamashi da kuma yanayin muhalli. Wannan yana da fa'ida musamman ga na'urori masu ƙarfin baturi, inda ƙarfin kuzari ke da mahimmanci.
4. Sauƙaƙan Ƙarfafawa da Matsala: Ƙananan LLMs sun fi sauƙi don aikawa da rarrabawa saboda ƙananan girman su. Ana iya haɗa su cikin aikace-aikace da tsarin daban-daban ba tare da na'urori na musamman ko manyan abubuwan more rayuwa ba. Wannan šaukuwa yana ba da damar ɗorewa mai faɗi kuma yana ba da damar haɓaka ƙarin ƙaƙƙarfan ƙa'idodi da tushen tushe.
5. Sirri da Mallakar Bayanai: Ta hanyar gudanar da ƙananan LLMs a gida, masu amfani za su iya kula da mafi girman iko akan bayanan su kuma rage buƙatar aika bayanai masu mahimmanci zuwa sabobin nesa ko dandamali na girgije. Wannan na iya taimakawa wajen magance matsalolin keɓantawa da bin ƙa'idodin kariyar bayanai.
6. Amfani da farashi: Ƙananan ƙira gabaɗaya suna buƙatar ƙarancin albarkatun lissafi, waɗanda zasu iya fassara zuwa ƙananan farashin aiki, musamman lokacin da ake aiki akan dandamalin girgije ko kayan aikin haya. Wannan kudin-tasiri na iya yin LLM fasaha ta fi dacewa ga ƙananan kungiyoyi da masu haɓaka mutum ɗaya.
7. Aikace-aikace na Musamman: Yayin da ƙananan ƙirar ƙila ba za su iya cimma matakin aiki ɗaya kamar manyan samfura a kan ayyuka na gaba ɗaya ba, ana iya daidaita su kuma an inganta su don takamaiman aikace-aikace ko yanki, mai yuwuwar fin ƙira mafi girma a waɗannan yankuna na musamman.

Yana da mahimmanci a lura cewa fa'idodin ƙananan LLMs suna zuwa tare da ciniki a cikin aiki da iyawa idan aka kwatanta da manyan takwarorinsu. Koyaya, ƙananan fa'idodin LLMs a cikin ingantaccen albarkatu, ɗaukar nauyi, da ingancin farashi na iya sanya su zaɓi mai tursasawa don aikace-aikace da yawa inda babban aikin aiki ba buƙatu bane mai mahimmanci.

Ƙananan LLMs Zaku Iya Gudu akan Na'urorin Gida

Rarraba

• Girma Model: Sigar tushe tana da kusan sigogi 66M, mafi ƙanƙanta fiye da sigogin BERT na 110M.
• description: DistilBERT sigar sigar BERT ce da aka ƙera, wanda aka ƙera don ƙarami da sauri yayin riƙe mafi yawan ayyukan BERT. Yana amfani da dabarun ƙwanƙwasa ilimi don damfara babban ƙirar BERT zuwa ƙaramin siga, yana mai da shi mafi inganci da sauƙin turawa akan na'urorin gida.
• Bukatun Hardware: Karamin girman DistilBERT yana ba shi damar aiki akan na'urori daban-daban na gida, gami da kwamfyutoci, tebur, har ma da manyan na'urorin hannu.

Rungumar Fuska Link: Rarraba

TinyBERT

• Girma Model: TinyBERT-4 yana da kusan 14M sigogi, yayin da TinyBERT-6 ke da kusan 67M.
• description: TinyBERT wani nau'in BERT ne mai karamci, wanda masu bincike a Jami'ar Carnegie Mellon da Google Brain suka kirkira. Yana amfani da ingantattun dabaru kamar na'urar-hikima da karkatar da hankali don cimma gagarumin matsawa samfurin yayin da ake ci gaba da yin gasa akan ayyukan NLP daban-daban.
• Bukatun Hardware: Ƙananan girman TinyBERT yana ba shi damar yin aiki akan nau'ikan na'urori masu yawa na gida, gami da kwamfyutoci marasa ƙarfi, na'urorin da aka saka, da na'urorin hannu.

Rungumar Fuska Link: TinyBERT

MobileBERT

• Girma Model: MobileBERT yana da kusan sigogi 25M, ƙarami sosai fiye da asalin tushen BERT.
• description: MobileBERT ƙaƙƙarfan ƙirar BERT ne mai inganci don wayar hannu da na'urori masu gefe. Yana amfani da dabaru kamar distillation ilmi da ƙididdigewa don rage girman samfurin yayin da yake riƙe babban aiki akan ayyuka masu yawa na NLP.
• Bukatun Hardware: Kamar yadda sunan ke nunawa, MobileBERT an inganta shi don aiki akan na'urorin hannu da sauran mahalli masu takurawa albarkatu.

Rungumar Fuska Link: MobileBERT

GASKIYA

• Girma Model: Ya bambanta dangane da tsari; ɗayan mafi ƙanƙanta shine tushe na ALBERT tare da yadudduka 12 da kawunan hankali 12.
• description: ALBERT (A Lite BERT) an tsara shi don ingantaccen amfani da ƙwaƙwalwar ajiya da saurin fahimta. Yana fasalta tsarin raba siga-Layer-Layer da rage girman abin sakawa. Yana da tasiri ga ayyuka daban-daban na NLP yayin da ya fi sauƙi fiye da ainihin BERT.
• Bukatun Hardware: Kyakkyawan ƙira na ALBERT yana ba shi damar aiki akan na'urori daban-daban na gida tare da matsakaicin ikon sarrafawa.

Rungumar Fuska Link: GASKIYA

GPT-2 Ƙananan

• Girma Model: GPT-2 Small yana da kusan sigogi 117M, ƙarami sosai fiye da manyan samfuran GPT-2.
• description: GPT-2 Small ƙaramin sigar mashahurin GPT-2 (Generative Pre-trained Transformer 2) wanda OpenAI ya haɓaka. Duk da yake ba ƙaramin ƙarfi ba kamar wasu samfuran, GPT-2 Small har yanzu yana da ɗan nauyi kuma ana iya amfani dashi don ayyuka kamar tsara rubutu, taƙaitawa, da ƙirar harshe.
• Bukatun Hardware: GPT-2 Ƙananan za a iya gudu akan kwamfutoci na sirri tare da matsakaicin ƙayyadaddun kayan aiki, kamar kwamfyutocin tsakiya ko kwamfutoci.

Rungumar Fuska Link: GPT-2 Ƙananan

DeciCoder-1B

• Girma Model: 1 biliyan sigogi
• description: DeciCoder-1B samfurin harshe ne da aka mayar da hankali kan tsara lamba da fahimta. Zai iya taimakawa tare da yin ƙididdigewa ayyuka kamar kammala lamba, fassarar tsakanin harsunan shirye-shirye, da bayanin lamba. An horar da shi a kan babban lambar tushe da bayanin harshe na halitta.
• Bukatun Hardware: Tare da ƙanƙantar girman sigina na biliyan 1, DeciCoder-1B na iya aiki akan na'urori daban-daban na gida kamar kwamfyutoci, tebur, da yuwuwar manyan na'urorin hannu ko kwamfutoci guda ɗaya.

Rungumar Fuska Link: DeciCoder - 1B

Phi-1.5

• Girma Model: 1.5 biliyan sigogi
• description: Phi-1.5 samfurin harshe ne na gaba ɗaya wanda zai iya samar da rubutu, amsa tambayoyi, da fahimtar harshe na halitta, da sauran ayyukan NLP. An ƙera shi don dacewa da yankuna da ayyuka daban-daban ta hanyar daidaitawa ko faɗakarwa.
• Bukatun Hardware: Phi-1.5's ƙaramin girman sigina biliyan 1.5 yana ba da damar a tura shi akan na'urori na gida tare da matsakaicin albarkatun kwamfuta, kamar kwamfyutocin kwamfyutoci, da na'urorin kwamfuta mai yuwuwar mafi girma ta hannu ko na'urorin kwamfuta guda ɗaya.

Rungumar Fuska Link: Phi-1.5

Dolly-v2-3b

• Girma Model: 3 biliyan sigogi
• descriptionDolly-v2-3b samfurin harshe ne mai bin umarni wanda ya ƙware wajen fahimta da aiwatar da dalla-dalla, faɗakarwar matakai da yawa da umarni cikin ayyuka daban-daban.
• Bukatun Hardware: Tare da sigogi biliyan 3, Dolly-v2-3b yana buƙatar na'urori na gida tare da matsakaicin matsakaicin ƙarfin kwamfuta, kamar manyan kwamfyutoci, tebur, ko wuraren aiki.

Rungumar Fuska Link: Dolly-v2-3b

StableLM-Zephyr-3B

• Girma Model: 3 biliyan sigogi
• description: StableLM-Zephyr-3B samfurin harshe ne da aka horar da shi don ba da amsoshi masu dogaro da gaskiya. An ƙera shi don zama tabbatacce kuma abin ƙira don ayyuka daban-daban na sarrafa harshe na halitta.
• Bukatun Hardware: Kamar Dolly-v2-3b, ma'auni na biliyan 3 StableLM-Zephyr-3B na iya aiki akan na'urorin gida tare da matsakaicin matsakaicin ƙarfin kwamfuta, kamar kwamfyutoci masu tsayi, tebur, ko wuraren aiki.

Rungumar Fuska Link: StableLM-Zephyr-3B

Farashin DeciLM-7B

• Girma Model: 7 biliyan sigogi
• descriptionDeciLM-7B samfurin harshe ne na gaba ɗaya don ayyukan sarrafa harshe daban-daban. Girman sigar sa na biliyan 7 yana ba da ingantacciyar aiki akan ƙananan ƙira yayin da har yanzu yana da ƙarfi don tura gida.
• Bukatun Hardware: Don gudanar da DeciLM-7B a cikin gida, masu amfani za su buƙaci samun damar yin amfani da tsarin tare da ƙarin kayan aiki masu ƙarfi, kamar manyan kwamfyutoci masu tsayi ko wuraren aiki tare da GPUs masu ƙarfi ko TPUs.

Rungumar Fuska Link: Farashin DeciLM-7B

Mistral-7B-Umarori-v0.2

• Girma Model: 7 biliyan sigogi
• description: Mistral-7B-Instruct-v0.2 samfurin harshe ne mai bin umarni wanda zai iya aiwatar da hadaddun umarni da ayyuka masu yawa.
• Bukatun Hardware: Mai kama da DeciLM-7B, Mistral-7B-Instruct-v0.2 yana buƙatar babban kayan aiki na gida, kamar kwamfutoci masu ƙarfi ko wuraren aiki, don gudanar da sigogin biliyan 7.

Rungumar Fuska Link: Mistral-7B-Umarori-v0.2

Orca-2-7B

• Girma Model: 7 biliyan sigogi
• description: Orca-2-7B samfurin harshe ne na buɗaɗɗen tushe wanda ke ba da amsoshi masu aminci, gaskiya, da haɗin kai. Yana da nufin samar da abubuwan da suka dace da kimar ɗan adam da xa'a.
• Bukatun HardwareMa'auni na biliyan 7 Orca-2-7B yana buƙatar kayan aikin gida masu ƙarfi kamar kwamfutoci masu inganci ko wuraren aiki don yin aiki yadda ya kamata.

Rungumar Fuska Link: Orca-2-7B

Amber

• Girma Model: 7 biliyan sigogi
• description: Amber samfurin harshe ne mai ɗawainiya da yawa da aka tsara don gudanar da ayyuka daban-daban na sarrafa harshe na halitta tare da babban aiki a cikin yankuna da aikace-aikace.
• Bukatun Hardware: Gudun ma'auni na biliyan 7 na Amber a cikin gida yana buƙatar samun dama ga babban kayan aiki, kamar kwamfutoci masu ƙarfi ko wuraren aiki tare da GPUs ko TPU masu ƙarfi.

Rungumar Fuska Link: Amber

BuɗeHathi-7B-Hi-v0.1-Base

• Girma Model: 7 biliyan sigogi
• description: OpenHathi-7B-Hi-v0.1-Base babban samfurin yaren Hindi ne, ɗaya daga cikin mafi girma da ake samu a fili don yaren Hindi. Yana iya fahimta da samar da rubutun Hindi.
• Bukatun Hardware: Kamar sauran nau'ikan 7B, OpenHathi-7B-Hi-v0.1-Base yana buƙatar babban aiki na gida, kamar kwamfutoci masu ƙarfi ko wuraren aiki, don yin aiki yadda ya kamata.

Rungumar Fuska Link: BuɗeHathi-7B-Hi-v0.1-Base

SOLAR-10.7B-v1.0

• Girma Model: 10.7 biliyan sigogi
• description: SOLAR-10.7B-v1.0 babban samfurin harshe ne na gabaɗaya yana tura iyakokin abin da zai iya gudana a cikin gida akan kayan masarufi. Yana ba da ingantaccen aiki don ayyuka daban-daban na NLP.
• Bukatun Hardware: Don ƙaddamar da SOLAR-10.7B-v1.0 a cikin gida, masu amfani za su buƙaci samun dama ga kayan aikin mabukaci na ƙarshe tare da GPUs masu ƙarfi ko saitin GPU masu yawa.

Rungumar Fuska Link: SOLAR-10.7B-v1.0

NexusRaven-V2-13B

• Girma Model: 13 biliyan sigogi
• description: NexusRaven-V2-13B babban samfurin harshe ne da aka mayar da hankali kan tsararrun rubutu mai ƙarewa a cikin yankuna da aikace-aikace daban-daban.
• Bukatun Hardware: A sigogin biliyan 13, NexusRaven-V2-13B yana buƙatar kayan aiki mai ƙarfi sosai, kamar manyan wuraren aiki ko saitin GPU da yawa, don gudanar da gida akan na'urorin masu amfani.

Rungumar Fuska Link: NexusRaven-V2-13B

Duk da yake waɗannan ƙananan LLMs suna ba da mahimmancin ɗaukar hoto da fa'idodin ingantaccen albarkatu, yana da mahimmanci a lura cewa ƙila ba za su cimma matakin aiki ɗaya ba kamar manyan takwarorinsu akan wasu hadaddun ayyuka na NLP. Koyaya, don aikace-aikacen da yawa waɗanda basa buƙatar aikin zamani, waɗannan ƙananan ƙirar zasu iya zama mafita mai amfani kuma mai sauƙi, musamman lokacin aiki akan na'urori na gida tare da ƙayyadaddun kayan ƙididdigewa.

Kammalawa

A ƙarshe, samun ƙananan ƙirar harshe waɗanda za su iya aiki a cikin gida akan na'urorinku alama ce ta ci gaba mai mahimmanci a AI da NLP. Waɗannan samfuran suna ba da ingantaccen haɗaɗɗen ƙarfi, inganci, da samun dama, suna ba ku damar aiwatar da ayyukan sarrafa harshe na yanayi na ci gaba ba tare da dogaro da sabis na girgije ko cibiyoyin bayanai masu ƙarfi ba. Yayin da kuke gwaji tare da waɗannan ƙananan LLMs, kuna buɗe sabbin hanyoyi don ƙirƙira da ƙirƙira a cikin ayyukanku, ko kun kasance ƙwararren mai haɓakawa, mai bincike, ko mai sha'awar sha'awa. The makomar AI baya iyakance ga manyan samfura; a maimakon haka, yana game da haɓaka yuwuwar kayan aikin da kuke da su. Gano abin da waɗannan ƙananan ƙira masu ƙarfi za su iya cimma muku!

Ina fatan kun sami wannan labarin yana da haske. Idan kuna da wasu shawarwari game da labarin, yi sharhi a ƙasa. Don ƙarin labarai, kuna iya komawa zuwa wannan mahada.

• Ƙunshin Ƙarfafa SEO & Rarraba PR. Samun Girma a Yau.
• PlatoData.Network Tsaye Generative Ai. Karfafa Kanka. Shiga Nan.
• PlatoAiStream. Web3 Mai hankali. Ilmi Ya Faru. Shiga Nan.
• PlatoESG. Carbon, CleanTech, Kuzari, Muhalli, Solar, Gudanar da Sharar gida. Shiga Nan.
• PlatoHealth. Biotech da Clinical Trials Intelligence. Shiga Nan.
• Source: https://www.analyticsvidhya.com/blog/2024/04/smallest-llms-that-you-can-run-on-local-devices/