# Isingeniso
Unjiniyela we-LLM akayona into efanayo nonjiniyela ojwayelekile wokufunda ngomshini. Lapho unjiniyela wokufunda komshini engase achithe izinyanga eqeqesha inethiwekhi ye-neural kusukela ekuqaleni, isikhungo somsebenzi kanjiniyela we-LLM ekuzivumelaniseni nezimo, ukuhlela, nokunikeza amamodeli olimi amakhulu aqeqeshwe kusengaphambili (ama-LLM). Umsebenzi ukuthatha imodeli yesisekelo enekhono futhi uyiguqule ibe into eyenza umsebenzi owusizo ngokuthembekile ngaphakathi komkhiqizo wangempela.
Isidingo sale ndima sikhule kakhulu ngo-2026. Izici ze-LLM ezisebenzise u-2023 no-2024 njengamademo angaphakathi manje sezithunyelwa njengezinhlelo zokukhiqiza, futhi izinhlangano zidinga onjiniyela abangazakha futhi bazinakekele. Amakhono ahilelekile acacile ngokwanele ukuthi isizinda sokufunda somshini esijwayelekile sikufikisa emgqeni wokuqala kodwa singaqhubeki kakhulu.
Lo mhlahlandlela uhlanganisa izindawo zamakhono ezinhlanu ngokulandelana kwazo: izisekelo, ukwaziswa kanye nokubiza kwamathuluzi, ukubuyisa, ukulungisa kahle nokuqondanisa, nokuphakela nokusebenza. Isinyathelo ngasinye siphetha ngephrojekthi ekhonkolo ongayivula umhleli bese uqala ukwakha namuhla. Ekugcineni, uzoba nesithombe esicacile sokuthi yini okufanele uyifunde nokuthi ungalandelani kanjani.
# Isinyathelo 1: Ukwakha Isisekelo
Uma usuvele usebenza ePython futhi unokuqonda okusebenzayo kokufunda komshini, ungadlula kulesi sinyathelo ngokushesha. Okubalulekile lapha ukwakha intuition mayelana nendlela ama-LLM aziphatha ngayo ezingeni lamathokheni, hhayi ukuphinde athole ukunakwa emigomeni yokuqala yezibalo.
Udinga ukuqonda kwezinga lokusebenza kwemiqondo emine: amathokheni (amamodeli amayunithi empeleni acubungula), ukushumeka (indlela amathokheni aba ngayo ama-vector endaweni enobukhulu obuphakeme), ukunaka (indlela imodeli ekala ngayo ubudlelwano phakathi kwamathokheni), kanye nebhulokhi ye-transformer njengeyunithi yezakhiwo ephindaphindayo. Awudingi ukwenza lokhu kusukela ekuqaleni. Udinga ukuwaqonda kahle ngokwanele ukuze ucabange ngokuthi kungani imodeli iziphatha ngendlela eziphatha ngayo.
I-PyTorch kanye ne Ubuso Obugonayo I-ecosystem (ikakhulukazi i-Transformers kanye ne-Datasets) iyindawo yokusebenza ezenzakalelayo yale ndima. Ukujwayelana kokubili kulindeleke.
Iphrojekthi: Layisha imodeli encane evuliwe usebenzisa ilabhulali ye-Transformers futhi uqalise ukukhiqizwa kombhalo kusuka ekwazisweni.
from transformers import AutoTokenizer, AutoModelForCausalLM
import torch
model_id = "HuggingFaceTB/SmolLM2-135M-Instruct"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id)
inputs = tokenizer("Explain what a transformer is:", return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=80)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))Lokhu kukunikeza umuzwa ophathekayo weluphu ye-tokenize-forward-decode ngaphambi kokuthi ubeke noma yini ngaphezulu kwayo.
# Isinyathelo sesi-2: Ukuklama Ukwaziswa kanye Nezinhlelo Zokwakha Zamathuluzi Okubiza
Ukwazisa akulona ikhono elithambile. Kuyi-lever yokuqala unjiniyela we-LLM ayifinyelelayo, futhi ukuyilungisa kudinga ukucabanga okuhlelekile: imilayezo yesistimu ehlelekile, izibonelo ezimbalwa ezifakwe ngamabomu, kanye nezikimu zokuphuma kwe-JSON eziphoqelela ukuziphatha kwemodeli kokuthile okungahlaziywa ngokuthembekile isistimu yomfula.
I-ceiling ibaluleke kakhulu njengephansi. Ukwazisa kukodwa kuyeka ukwanele uma udinga imodeli ukuze wenze ngokwesimo sangaphandle kunokuba nje ucabange ngombhalo. Kulapho ukushaya kwamathuluzi kungena khona, futhi ngo-2026 kuyikhono lesigaba sokuqala kuwo wonke amamodeli amakhulu we-API, hhayi iqhinga elithuthukile.
Ukushaya kwethuluzi kusebenza ngokunikeza imodeli isethi yamasiginesha omsebenzi futhi uyivumele inqume ukuthi izocela ini ngokusekelwe esicelweni somsebenzisi. Imodeli ibuyisela ucingo oluhlelekile; ikhodi yakho iyayenza futhi ibuyisela umphumela; imodeli ihlanganisa lowo mphumela empendulweni yayo elandelayo. Le loop iyimbewu yezakhiwo zesistimu ye-agent, ozoyinweba esinyathelweni sesi-3.
Isiqondisindlela esisodwa okufanele ukwazi ngaso: uma usunamamethrikhi okuhlola ongawasebenzisa ngokumelene, nezinhlaka zokuthuthukisa ngokushesha zohlelo ezifana I-DSPy ikuvumela ukuthi uphathe ukwakhiwa okusheshayo njengenkinga yokuthuthukisa kunokuphatha umsebenzi wokushuna mathupha.
Iphrojekthi: Ithuluzi lomugqa womyalo eliphendula umbuzo womsebenzisi ngokushayela isimo sezulu sangaphandle noma i-API yesitoko ngokushaya kwethuluzi lomdabu, bese lifometha impendulo.
tools = [
{
"name": "get_weather",
"description": "Get current weather for a city",
"input_schema": {
"type": "object",
"properties": {"city": {"type": "string"}},
"required": ["city"]
}
}
]
response = client.messages.create(
model="claude-sonnet-4-20250514",
max_tokens=512,
tools=tools,
messages=[{"role": "user", "content": "What is the weather in Bangkok?"}]
)Imodeli ibuyisela a tool_use block okuqukethwe. Ikhodi yakho iphatha ukuthunyelwa, ishayela i-API yangempela, futhi iwufunze umphumela.
# Isinyathelo sesi-3: Ukwakha Izinhlelo Zokubuyisa Ngalé Kwezisekelo
I-Retrieval-augmented generation (RAG) manje isiyisakhiwo esijwayelekile sezinhlelo zokusebenza ze-LLM ezidinga ukuphendula imibuzo ngedatha eyimfihlo noma ebuyekezwa njalo. Ngaphambi kokwakha noma yini ethuthukisiwe, khululeka ngepayipi lesisekelo: hlukanisa amadokhumenti abe amasegimenti, shumeka ingxenye ngayinye ku-vector, gcina ama-vector kusizindalwazi se-vector, thola izingcezu ezifanele kakhulu ngesikhathi sombuzo, bese uwahlanganisela efasiteleni lomongo wemodeli.
Ubunjiniyela bangempela buqala lapho ukubuyisa okungazi lutho sekusebenza. Ukusesha kwamagama angukhiye angacacile kanye nokusesha okushumekiwe okuminyene ngakunye kugeja imibuzo ehlukene. Ukuwahlanganisa njengosesho oluxubile, bese usebenzisa i-reranker ukuze uhlele kabusha imiphumela ngokuhambisana nombuzo othile, kuphakamisa ngokuthembekile ukunemba kokubuyisa kumadokhumenti angempela. Umzila we-Semantic, lapho isihlukanisi sithumela imibuzo emthonjeni ofanele ngaphambi kokuthi kuqale ukubuyisa, siphatha amasistimu emithombo eminingi ngaphandle kokululaza kunoma iyiphi eyodwa.
Izindlela zokwehluleka ezivamile: izingcezu ezinkulu kakhulu zesignali yokuhlambulula, izingcezu ezincane kakhulu zilahlekelwa umongo, kanye nokugeja kokubuyisa kukhiqiza izimpendulo ezingalungile ezizwakala ngokuzethemba. Udinga ukukala ikhwalithi yokubuyisa ngokwehlukana nekhwalithi yokukhiqiza ukuze ulungise lokhu.
Gcina uchungechunge lwe-agency olusuka esinyathelweni sesi-2 lapha engqondweni: ukubuyisa ithuluzi umenzeli angalibiza, ekhetha ukuthi uzobheka nini okuthile ngokusekelwe embuzweni. Ngedatha eyimfihlo eyinkimbinkimbi enobudlelwano bebhizinisi obuminyene, izindlela zegrafu yolwazi (ngezinye izikhathi zibizwa ngokuthi i-GraphRAG) zinikeza inketho ejulile eyisisekelo okufanele ihlolwe.
Izinketho zesitolo seVector zisukela endaweni (FAISS, I-Chroma) ukuphatha (Weaviate, Iphayini). I-LangChain, I-LlamaIndexfuthi I-LangGraph yizinhlaka eziyinhloko ze-orchestration.
Iphrojekthi: Isistimu yokuphendula idokhumenti esebenzisa ukuziphendulela ukuze ibhale kabusha umbuzo lapho umzamo wokuqala wokubuyisa ubuyisela imiphumela yokungazethembi okuphansi.
from langchain_community.vectorstores import Chroma
from langchain_openai import OpenAIEmbeddings
embedder = OpenAIEmbeddings()
vectorstore = Chroma.from_documents(docs, embedder)
retriever = vectorstore.as_retriever(search_kwargs={"k": 5})
results = retriever.invoke("What are the contract renewal terms?")Ngemva kokubuyisa, bhala amaphuzu. Uma ukuzethemba kungaphansi komkhawulo, bhala kabusha umbuzo ngemodeli bese uthola futhi ngaphambi kokukhiqiza.
# Isinyathelo sesi-4: Ukushuna Kahle Nokuqondanisa Amamodeli
Ukwaziswa nokubuyisa kuxazulula izinkinga eziningi. Ukushuna kahle kufanelekile uma udinga imodeli ukuze yamukele ngokungaguquki ifomethi ethile, ithoni, noma ulwazimagama lwesizinda ukwaziswa okungakwazi ukukuphoqelela ngokwethembeka, noma lapho udinga ukwehlisa izindleko ze-inference ngokuziphatha kwe-distilling ube yimodeli encane.
Izindlela ezisebenza kahle ngepharamitha ziyisiqalo esijwayelekile. I-Low-Rank Adaptation (LoRA) kanye nokwahluka kwayo okulinganiselwe kwe-QLoRA kukuvumela ukuthi uqeqeshe isethi encane yezisindo ze-adaptha phezu kwemodeli yesisekelo esiqandisiwe, uzuze ushintsho olukhulu lokuziphatha ngengxenyana yezindleko zokubala zokulungisa kahle. I I-PEFT futhi I-TRL amalabhulali ku-Hugging Face ecosystem aphatha kokubili.
I-Direct Preference Optimization (DPO) manje isiyindlela evamile yokuvumelanisa ukuziphatha okuyimodeli emiphumeleni ekhethwayo ngaphandle kwenkimbinkimbi yokufunda okuqiniswayo okuvela kumpendulo yomuntu (RLHF). Isebenza ngamapheya okuphothula okuthandwayo nokunqatshiwe futhi ithathe indawo yezindlela ezisuselwe ku-PPO zokuqondanisa ithoni nesitayela.
Ukukhethwa kwesethi yedatha yilapho isikhathi esiningi sobunjiniyela siya khona. Imodeli ecushwe kahle inhle kuphela njengezibonelo zayo zokuqeqesha, futhi ukwakha amapheya ahlanzekile, amele okuthandwayo kuthatha isikhathi eside kunokuqeqeshwa okwenziwa ngokwako.
Ukuhlola kuwumsebenzi wobunjiniyela besigaba sokuqala lapha: ukwakha amasethi e-eval yohlelo, ukubhala amasudi okuhlola ahlola ifomethi yokuphumayo nokubambelela kweqiniso, nokusebenzisa ama- guardrail abamba izindlela zokuhluleka ngaphambi kokuba afinyelele kubasebenzisi. I-Ragas futhi Phoenix zingamathuluzi asebenzayo akho kokubili ukuhlola nokubonakala.
Iphrojekthi: Lungisa kahle imodeli encane evulekile ukuze ifane nethoni ethile yenkampani, bese ukala ukubambelela uqhathanisa nesisekelo usebenzisa umhloli ohleliwe.
from peft import LoraConfig, get_peft_model
from transformers import AutoModelForCausalLM
base_model = AutoModelForCausalLM.from_pretrained("HuggingFaceTB/SmolLM2-360M")
lora_config = LoraConfig(r=16, lora_alpha=32, target_modules=["q_proj", "v_proj"])
model = get_peft_model(base_model, lora_config)
model.print_trainable_parameters()Okukhiphayo kuzobonisa cishe u-1–2% wamapharamitha esewonke amakwe njengangaqeqesheka, okuyisici sokucushwa kwe-LoRA okuphumelelayo.
# Isinyathelo sesi-5: Ukunikeza kanye nokusebenzisa izinhlelo zokusebenza ze-LLM
Ukuthola imodeli esebenza endaweni futhi uyithole inikezela ngethrafikhi yokukhiqiza izinkinga ezahlukene zobunjiniyela. Amamodeli ezisindo ezivulekile adinga ingqalasizinda ye-inference ephethe i-batching (ukukhonza izicelo eziningi ngesikhathi esisodwa ukuze kwandiswe ukusetshenziswa kwe-GPU) kanye nokulinganisa (ukunciphisa ukunemba kwezinombolo ukuze kwehliswe inkumbulo yonyawo nokwandisa ukuphuma). i-vLLM iwukukhetha okujwayelekile kokuphakelwa okuthuthukisiwe; U-Ollama iphatha ukuthuthukiswa nokuhlola kwendawo. amabhayithithi ihlanganisa i-4-bit kanye ne-8-bit quantization.
I-LLMOps isendlalelo sokusebenza: ukulandelela ukusetshenziswa kwethokheni ngesicelo ngasinye, okokufaka kokungena kanye nemiphumela yokulungisa iphutha nokuhambisana, ukwaziswa kwenguqulo okuhambisana nekhodi yohlelo lokusebenza ukuze ukwazi ukukhiqiza kabusha noma yikuphi ukuziphatha okudlule, kanye nezindleko zokuqapha nokubambezeleka ngokuhamba kwesikhathi. Lezi yizinqubo ezihlukanisa i-prototype esebenzayo ohlelweni lokukhiqiza olugcinekayo. Izisindo & Ukuchema iphatha ukulandelela ukuhlolwa; I-Phoenix ihlanganisa ukubonwa kokukhiqiza.
Gcina lo msebenzi kusendlalelo sohlelo lokusebenza. Okugxilwe kukho lapha ukuthembeka nephrofayili yezindleko zohlelo lwakho lokusebenza kanye nesisekelo sayo sekhodi, hhayi ukwakheka kwengqalasizinda ebanzi yenhlangano.
Iphrojekthi: Goqa isistimu yokubuyisa kusukela kusinyathelo sesi-3 ngemuva kwe-API engasindi bese wengeza iloga ye-telemetry elandelela ukubala kwamathokheni, ukubambezeleka, kanye nezindleko ezilinganiselwe ngekholi ngayinye.
from fastapi import FastAPI
import time
app = FastAPI()
@app.post("/query")
async def query_endpoint(question: str):
start = time.time()
response = rag_chain.invoke(question)
latency_ms = (time.time() - start) * 1000
log_telemetry(question, response, latency_ms)
return {"answer": response, "latency_ms": latency_ms}Ukwengeza i-telemetry ehlelekile ngaphambi kwesikhathi kukhokha izinzuzo: izimanga zezindleko kanye nokwehla kwesikhashana kulula kakhulu ukukubamba uma unedatha yesisekelo.
# Izinsiza Zokufunda Ezinconyiwe
Izifundo kanye nezifundo:
Izincwadi:
- Amamodeli Olimi Olukhulu Wokusebenzisa izandla ngu-Jay Alammar kanye no-Maarten Grootendorst
- Yakha Imodeli Yolimi Elikhulu (Kusuka Ekuqaleni) nguSebastian Raschka
Amadokhumenti afanele ukubekisa: amadokhumenti e-Hugging Face PEFT, okokufundisa kwe-LangGraph kuma-agent loops, kanye negayidi yokuphakela ye-vLLM.
# Imicabango yokugcina
Lezi zinyathelo ezinhlanu zakha isitaki lapho isendlalelo ngasinye sincike kulesi esingezansi. Izisekelo zikunikeza ulwazimagama lokubonisana mayelana nokuziphatha okuyimodeli. Ukwaziswa nokushaya ithuluzi kukunikeza isixhumi esibonakalayo esiyinhloko ukuze ukwazi imodeli. Ukubuyisa kuxhumanisa amamodeli nolwazi lwangaphandle. Ukushuna kahle nokuqondanisa kukuvumela ukuthi ulolonge kabusha imodeli yokuziphatha ngezidingo ezithile. Ukunikeza kanye nokusebenza kuguqula konke kube into esebenza ngokuthembekile ngaphansi komthwalo.
Umugqa wesikhathi ongokoqobo womuntu onesizinda sokufunda somshini izinyanga ezintathu kuya kweziyisithupha zomsebenzi ogxilisiwe ukuze wakhe ukuzethemba kuzo zonke izindawo ezinhlanu, iphrojekthi yokuqala ihanjiswe kahle ngaphambi kwalokho. Iphothifoliyo ibaluleke ngaphezu kwezitifiketi kule ndima. Idemo esesidlangalaleni yesistimu yokubuyisa esebenzayo noma imodeli ecushwe kahle enemiphumela yokuhlola ebhaliwe ibonisa ikhono ngokuqondile kunanoma yisiphi isifundo.
Uma intshisekelo yakho idonsela ekwakhiweni kwesistimu, ingqalasizinda, kanye nesakhiwo senhlangano esikhundleni sokwakha ezingeni lekhodi, indlela ehambisanayo okufanele uyihlole wumsebenzi womklami we-AI. Lezi zindima ezimbili zabelana ngezisekelo kodwa zehluka kakhulu ngemva kwesinyathelo soku-1.
Qala ngesinyathelo soku-1 kuphela uma usidinga. Bese uthumela into encane ekupheleni ngaphambi kokungena ujule kunoma iyiphi indawo eyodwa.
Vinod Chugani unguthisha we-AI kanye nesayensi yedatha ovala igebe phakathi kobuchwepheshe be-AI obusafufusa kanye nokusebenzisa okusebenzayo kochwepheshe abasebenzayo. Izindawo agxile kuzo zifaka i-agent AI, izinhlelo zokusebenza zokufunda ngomshini, nokugeleza komsebenzi okuzenzakalelayo. Ngomsebenzi wakhe njengomeluleki nomfundisi wezobuchwepheshe, uVinod uye wasekela ochwepheshe bedatha ngokuthuthukiswa kwamakhono kanye noshintsho lomsebenzi. Uletha ubungcweti bokuhlaziya kusukela kwezezimali zenani kuya endleleni yakhe yokufundisa yokufundisa. Okuqukethwe kwakhe kugcizelela amasu nezinhlaka ezingasetshenzwa ochwepheshe abangazisebenzisa ngokushesha.
