Uma ukhetha ukuthi imibhalo yakho ye-PDF, amarisidi noma imininingwane yomuntu siqu ingaweli ezandleni zezinkampani zangaphandle njenge-OpenAI, iMicrosoft, i-Google nezinye, uzokujabulela ukwazi ukuthi ungacubungula futhi ufunde ama-PDF kukhompyutha yakho noma kukhompyutha yakho. inethiwekhi yakho yomuntu siqu noma yangasese sibonga imodeli ye-Mistral AI. Uzojabula ukwazi ukuthi ungacubungula futhi ufunde ama-PDF kukhompuyutha yakho yomuntu siqu noma eyimfihlo noma inethiwekhi usebenzisa imodeli ye-Mistral AI.
Kulezi zinyanga eziyi-18 noma ngaphezulu ezedlule, ubuhlakani bokwenziwa (AI) bubone intuthuko enkulu, ikakhulukazi endaweni yokucubungula amadokhumenti, ngenxa yamamodeli amakhulu ezilimi akwazi ukufunda. Enye intuthuko enjalo ukusetshenziswa kwe-AI ukufunda nokucubungula imibhalo ye-PDF endaweni. Lo mhlahlandlela uchaza kabanzi ukuthi ungawagcina kanjani amadokhumenti akho e-PDF evikelekile ngokuwacubungula kukhompuyutha yakho noma kunethiwekhi yangakini. Kusetshenziswa umtapo ovulekile we-Katana ML ukucubungula imibhalo ye-PDF endaweni ngemodeli ye-Mistral AI.
“I-Mistral-7B-v0.1 imodeli encane kodwa enamandla elingana namacala amaningi okusebenzisa. I-Mistral 7B ingcono kune-Llama 2 13B kuwo wonke amabhentshimakhi, inamandla emvelo okubhala amakhodi nobude bokulandelana obungu-8k. Ikhishwe ngaphansi kwelayisensi ye-Apache 2.0, futhi sikwenze kwaba lula ukuyisebenzisa kunoma yiliphi ifu. "
I-Katana ML iwuhlaka oluvulekile lwe-MLOps olungasetshenziswa emafini noma emagcekeni. Inikeza ama-API wokufunda womshini asezingeni eliphezulu abhekana nezinhlobonhlobo zamacala okusetshenziswa. Enye yalezi zinhlelo zokusebenza icubungula imibhalo ye-PDF kusetshenziswa imodeli ye-Mistral 7B. Le modeli, naphezu kobukhulu bayo obuncane, ibonisa ukusebenza okumangazayo nokuzivumelanisa nezimo.
Ungawafunda kanjani futhi uwacubungule kanjani ama-PDF endaweni usebenzisa i-Mistral AI?
I-Mistral 7B iyimodeli yepharamitha eyizigidi eziyizinkulungwane ezingu-7,3 edlula ozakwabo, i-Llama 2 13B ne-Llama 1 34B, kumabhentshimakhi ahlukahlukene. Ize isondele ekusebenzeni kweCodeLlama 7B ekubhaleni amakhodi, kuyilapho igcina ubuhlakani bayo emisebenzini yesiNgisi. Imodeli isebenzisa indlela ye-GQA (Ukunakwa kwemibuzo eyiqembu) ukuze kufinyelelwe ngokushesha kanye nendlela ye-SWA (Sliding Window Attention) ukucubungula ukulandelana okude ngezindleko eziphansi. Isifanekiso sikhishwe ngaphansi kwelayisensi ye-Apache 2.0 futhi singasetshenziswa ngaphandle kwemikhawulo.
Inqubo yokusebenzisa le modeli ukufunda nokucubungula ama-PDF endaweni ingaqhutshwa ezinkundleni ezifana ne-Google Colab noma emshinini wasendaweni. Ukukhetha phakathi kwakho kokubili kuncike kulokho okuthandwa ngumsebenzisi kanye nezidingo. I-Google Colab inikeza inzuzo yokucubungula okusekelwe efwini, isusa isidingo sezingxenyekazi zekhompuyutha ezisezingeni eliphezulu. Kodwa-ke, futhi inemikhawulo, njengenombolo ekhawulelwe yokusetshenziswa kwamahhala kwe-GPU. Ngakolunye uhlangothi, ukusebenzisa umshini wendawo kuvumela ukulawula okukhulu nokwenza ngokwezifiso. Nokho, isivinini sokucubungula singahamba kancane ngenxa yemikhawulo yezingxenyekazi zekhompuyutha.
Ukukhombisa inqubo, ake sithathe isibonelo se-invoyisi ngefomethi ye-PDF. Isinyathelo sokuqala ukuhlanganisa inqolobane ye-Katana ML futhi ufake izimfuneko ezidingekayo. Umsebenzisi ube eselanda imodeli elinganiselwe ngokusekelwe kumthamo we-RAM wesistimu. Ifayela lokumisa libe selihlelwa ukuze kuthuthukiswe isivinini nekhwalithi. Idatha evela ku-PDF iguqulwa ibe okokushumeka futhi igcinwe ku-Vector DB, inqubo eyaziwa ngokuthi umjovo wedatha. Ifayela le-main.py libe selisetshenziswa ukuze libuze imibuzo futhi lithole izimpendulo ngokusekelwe kudatha ecutshunguliwe.
Naphezu kwamakhono ayo amangalisayo, imodeli ye-Mistral AI inemikhawulo yayo. Isivinini sokucubungula singase sihambe kancane ngenxa yemikhawulo yobuchwepheshe bamanje. Ngaphezu kwalokho, njenganoma iyiphi imodeli ye-AI, i-Mistral 7B ayigonyiwe “ekubonweni kwemibono” noma amaphutha. Lezi yizimo lapho i-AI ikhiqiza izimpendulo ezingalungile noma ezingenangqondo.
Nokho, ukusetshenziswa okungenzeka kwalobu buchwepheshe kukhulu. Isibonelo, ingasetshenziswa ukukhipha ulwazi oluhlelekile kumadokhumenti angahlelekile, njengama-invoyisi noma izinkontileka. Lokhu kungalula kakhulu izinqubo ezimbonini ezifana nezezimali, umthetho kanye nokuphatha.
Uma ubheka phambili, kunamathuba amaningana okuthuthukisa nokuthuthukisa. Isibonelo, ukushuna okwengeziwe kwemodeli kungathuthukisa ukusebenza kwayo. Ukwengeza, ukuthuthuka kobuchwepheshe behadiwe kungasheshisa kakhulu isikhathi sokucubungula.
Ukusebenzisa umtapo ovulekile we-Katana ML ukucubungula amadokhumenti e-PDF endaweni ngemodeli ye-Mistral ye-AI kuwuhlelo oluthembisayo lobuchwepheshe be-AI. Naphezu kwemikhawulo yayo yamanje, inikeza ukuqonda ngekusasa lokucutshungulwa kwemibhalo kanye namandla e-AI okuguqula imisebenzi evamile ibe yizinqubo ezizenzakalelayo.
Funda kabanzi Umhlahlandlela:
- Qalisa i-Llama 2 Engahloliwe kanye namanye ama-LLM endaweni usebenzisa i-Ollama ukuze uthole ubumfihlo nokuphepha
- Faka kalula amamodeli e-AI angokwezifiso endaweni eno-Ollama
- Imodeli entsha ye-Mistral Large AI idlula i-GPT-3.5 ne-Llama2-70B
- Isebenzisa i-Llama 2 endaweni ku-Apple M3 Silicon Macs
- Ungayifaka kanjani i-Ollama endaweni ukuze usebenzise i-llama 2, ikhodi ye-llama namanye amamodeli we-LLM?
- Uchwepheshe we-Mistral AI Mixtral 8x7B ohlanganisa imodeli yobuhlakani bokwenziwa kuye kwavezwa amabhentshimakhi ahlaba umxhwele
