AI4Kids — ai4kids.courses Online School← Sab Free Apps
AI4Kids · Class 5 se 10

LLM ki Lafz Machine

ChatGPT, Claude aur Gemini jaise AI ke peeche ek LLM hota hai. Aaj hum khud khel kar dekhenge ke yeh machine likhti kaise hai.

Meri class:
LLM = Large Language Model

Iska sab se bara kaam sirf ek hai: agla lafz guess karna. Bas yehi kaam bar bar, bohat tezi se. Isi se poore jawab, kahaniyan aur code ban jaate hain.

1

Agla lafz kya hoga?

Neeche wala jumla parho. LLM har mumkin lafz ko ek chance (probability) deta hai. Phir unmein se ek chunta hai.

Thanda — har baar same jawabGaram — ajeeb aur naye jawab

Button dabao aur dekho LLM kaunsa lafz chunta hai.

Gehrai mein (Class 8–10): Temperature asal mein probabilities ko "tez" ya "naram" karta hai. Kam temperature par sab se bara chance aur bara ho jata hai. Zyada temperature par chhote chances bhi barh jate hain. Is liye coding ke liye kam aur kahani likhne ke liye zyada temperature acha hota hai.
2

Tokens: LLM lafz kaise kaat-ta hai

LLM poore lafz nahi parhta. Woh text ko chhote tukron mein kaat-ta hai jinhein tokens kehte hain. Jaise roti ke niwaale!

Tokens: 0

Yeh ek asaan simulation hai. Asli LLM ka tokenizer thora mukhtalif kaat-ta hai, lekin idea yehi hai.

Gehrai mein: Har token ko ek number (ID) milta hai. LLM andar sirf numbers ke saath math karta hai. Aam English lafz aksar 1 token hote hain, jabke Roman Urdu ke lafz zyada tukron mein toot-te hain kyun ke model ne unhein kam dekha hota hai.
3

Apna chhota LLM train karo

Asal LLM ne internet ki arbon (billions) lines parhi hain. Hamara mini model sirf neeche wali lines se seekhega: "is lafz ke baad aksar kaunsa lafz aata hai?"

Training data

Ab model se likhwao

…

Kai baar "Likho!" dabao. Har baar thora mukhtalif jumla banega. Model ne jo parha, wahi dohraata hai — naye lafz khud nahi bana sakta.

Gehrai mein: Hamara model sirf pichla 1 lafz dekhta hai (isay bigram kehte hain). Asli LLM "Transformer" naam ka design use karta hai jo pichle hazaron tokens ek saath dekh sakta hai. Is ko "attention" kehte hain — model khud tay karta hai ke kis lafz par zyada dhyan dena hai.
4

Kya LLM hamesha sach bolta hai?

Nahi! LLM sirf achha lagne wala agla lafz chunta hai, zaroori nahi woh sach ho. Ghalat baat ko poore yaqeen se kehna hallucination kehlata hai. Har jawab ko parkho:

5

Quiz time!

Aaj humne kya seekha

Agla lafzLLM har baar sirf agla lafz guess karta hai.
TokensText chhote tukron mein kat kar numbers banta hai.
TrainingModel wahi seekhta hai jo usne parha.
HallucinationLLM ghalat ho sakta hai — hamesha check karo.

Teacher ke liye tips

  • Station 1 par class se pehle guess karwao, phir button dabao.
  • Station 3 mein har bacha apni ek line add kare, phir model ka jawab dekhein.
  • Class 8–10 ke liye "Gehrai mein" boxes zor se parhwao. Class 5–7 button dabane se yeh chhup jaate hain.
  • Homework: ChatGPT ya Claude se ek sawal poocho aur book se uska jawab check karo.