Lesson · Hinglish · ⏱ 7 min padhne mein

💬 LLM kya hai? Tokens, context aur temperature

🏠 Asli zindagi ki misaal

Typing karte waqt phone ka keyboard agla word suggest karta hai. LLM (large language model) wahi idea hai, lekin bahut bada: internet ke bahut bade hisse ke text pe train hokar woh agla word itni acchi tarah predict karta hai ki poore essays, code aur jawaab ban jaate hain. 'Context window' yeh hai ki woh ek baar mein conversation ka kitna hissa yaad rakh sakta hai, jaise whiteboard ka size.

💡 Aasaan shabdon mein

LLMs tokens mein padhte aur likhte hain: text ke tukde, English mein lagbhag ¾ word. Hindi aur doosri languages mein aksar ek word ke zyaada tokens lagte hain. Aap tokens ke hisaab se pay karte ho, aur limit bhi tokens ki hi hoti hai. Context window maximum tokens hain (aapka prompt, documents aur jawaab milakar) jo model ek baar mein handle kar sakta hai. Temperature randomness control karta hai: low (0–0.3) focused, repeatable jawaab deta hai, aur zyaada value zyaada creative hoti hai. Jab tak aap tools ya documents na do, LLMs kuch 'look up' nahi karte. Woh likely text generate karte hain, isliye confidently galat bhi ho sakte hain (hallucinate).

🎬 Dekho kaise kaam karta hai

LLM har possible agle token ko score karta hai, ek chunta hai (yahan 'ginger'), use text mein jodta hai, aur repeat karta hai, ek-ek token karke.

🪜 Step by step

Aapka text tokens (shabdon ke tukde) mein toota jaata hai. "Unbelievable" ban sakta hai "un", "believ", "able".
Model ek saath jo kuch dekhta hai, aapke instructions, documents, chat history aur uska khud ka jawaab, sab context window mein fit hona chahiye.
Model har mumkin agle token ki probability nikaalta hai, ek chunta hai, jodta hai, aur repeat karta hai.
Temperature tay karta hai woh chunaav kitna adventurous ho: facts aur formats ke liye kam, brainstorming ke liye zyaada.
API calls ke beech isko kuch yaad nahi rehta. Chat apps har baar poori baatcheet dobara bhejti hain, isliye lambi chats mehengi padti hain.

🔑 Zaroori shabd

Token: Woh unit jisse LLM padhta hai aur billing hoti hai, word ka tukda.
Context window: Maximum text jo model ek baar mein dhyaan mein rakh sakta hai.
Hallucination: Bharose se bola gaya jawaab jo banaaya hua ya galat hai.

👀 Example dekho

text

Prompt: "Suggest a name for my chai startup"
temperature 0.0 → "Chai Point" (same every time)
temperature 1.0 → "Kettle & Kulhad", "Brew Bandhu", "Sip Sutra"… (varies)

Tokens: "Suggest a name for my chai startup" ≈ 8 tokens

Facts aur code ke liye low temperature, brainstorming ke liye zyaada.

⚠️ Aam galtiyaan jinse bachna hai

Jawaabon ko facts maan lena. LLMs sahi lagne wala text banate hain; zaroori cheezein verify karo ya sources se ground karo.
Knowledge cutoff ignore karna. Model training data ke baad ki baatein nahi jaanta jab tak aap woh jaankari na do.

✅ Jaldi se check karo

LLM galat fact bhi bharose se kyun bol sakta hai?

Woh jaan-boojhkar jhooth bol raha hai
Woh likely-sounding text banata hai aur khud facts check nahi karta
Uska internet slow hai

Jawaab dekho

B. Woh likely-sounding text banata hai aur khud facts check nahi karta: LLMs plausible text predict karte hain. Retrieval ya tools ke bina unke paas fact-check ka koi step nahi hota.

🙋 Seekhne wale yeh poochhte hain

Ek page text mein kitne tokens?

English ke ek page mein lagbhag 500 se 700 tokens. Hindi aur doosri scripts aksar same matlab ke liye zyaada tokens leti hain.

Same prompt alag jawaab kyun deta hai?

Kyunki token probabilities se sample hota hai. Zyaada consistent output ke liye temperature kam karo (0 ki taraf).

📌 Yaad rakho

LLM = super-powered next-word predictor. Tokens, context aur temperature ka dhyaan rakho.

📖 Is lesson ke shabd (AI dictionary)

Poora course, aapke roadmap ke hisaab se

Progress tracking, quick checks aur step-by-step order ke saath. Yeh lesson in courses ka hissa hai:

AI Foundations Path · AI Product Engineer · LLM / Agent Engineer · AI Forward Deployed Engineer · AI Product Manager

Course mein seekho