مین مواد پر جائیں

Yeh source assessment Module 6 (Generative AI) aur Module 9 (Agentic AI) dono ke liye hai.

Module 9 Test: Agentic AI

Ye tumhara progress tracker hai. 20 sawal: 5 Easy, 10 Moderate, 5 Hard . Pehle sab sawal khud hal karo (code ko Colab me chalane se pehle apne dimagh se output sochna), phir Answer Key se check karo.

ScoreMatlab
18 se 20Zabardast! Modules 6 aur 9 par tumhari grip mazboot hai.
14 se 17Bohat acha! Kamzor chapter dobara parh lo.
10 se 13Theek hai. Code khud chala kar dobara dekho.
10 se kamFikr nahi. Chapters 29 se 33 dobara parho aur code chala kar dekho.
EASY​
Q1. Transformer ne RNN/LSTM ke muqable me sab se bari tabdeeli kya ki?​
  • A) Text ko sirf ek ek lafz karke parha

  • B) Sirf images par kaam kiya

C) Tokens ko ek sath (parallel) process kiya aur attention se rishte pakre

  • D) Training ki zaroorat hi khatam kar di
Q2. Attention me Query (Q) ka matlab kya hai?​
  • A) Main kya dhoond raha hoon (talash)

  • B) Token ka asal maal/maloomat

  • C) Token ki jagah (tartib)

  • D) Loss ki qeemat

Q3. LLM ki pre-training ka bunyadi maqsad kya hota hai?​
  • A) Tasveerein classify karna

  • B) Data ko saaf karna

  • C) Sirf translation karna

  • D) Agla token sahi andaza lagana

Q4. OpenAI API key ke bare me sahi baat kaun si hai?​
  • A) Code me likh kar GitHub par daal deni chahiye

  • B) Isay environment variable ya Colab Secrets me rakhna chahiye

  • C) Dosto ko WhatsApp par bhej deni chahiye

D) Website ke URL me likh deni chahiye

Q5. Function calling me asal function kaun chalata hai?​

A) Model khud

  • B) OpenAI ka server khud

  • C) Tumhara apna code

D) Browser

​
MODERATE​
Q6. Scaled dot-product attention me scores ko sqrt(d_k) se kyun bhaagte hain?​
  • A) Taake output chhota ho jaye

B) Taake softmax saturate na ho (ek hi weight 1 ke qareeb na pohanche)

  • C) Taake memory bache

D) Taake tokens ki tartib ka pata chale

Q7. Transformer me Positional Encoding ki zaroorat kyun parti hai?​
  • A) Loss kam karne ke liye

B) Vocabulary barhane ke liye

C) Heads ki ginti tay karne ke liye

D) Attention ko tokens ki tartib khud nahi pata hoti​
Q8. BERT aur GPT me sahi farq kaun sa hai?​

A) BERT encoder-only (dono taraf dekhta, masked LM); GPT decoder-only (sirf pichla, agla token)

  • B) BERT decoder-only aur GPT encoder-only hai

  • C) Dono bilkul ek jaise hain

D) BERT text generate karta hai, GPT sirf classification

Q9. Sampling me temperature ko 0.2 se 2.0 karne se kya hota hai?​

A) Jawab zyada pakka aur ek jaisa ho jata hai

  • B) Model ki zehaanat barhti hai

C) Jawab zyada random aur alag alag ho jate hain

D) Jawab ka size chhota ho jata hai

Q10. RAG ka sab se bara faida kya hai?​

A) Model ke weights badal jate hain

B) GPU ki zaroorat khatam ho jati hai

  • C) Tokens ka kharcha zero ho jata hai

D) Taaza/apne documents se jawab aur sources, jis se hallucination kam hota hai

Q11. Prompt me 3 misalein (input -> output) dikha kar model se naya kaam karwana kya kehlata hai?​
  • A) Fine-tuning

  • B) Few-shot prompting

C) Embedding

D) Tokenization

Q12. Fine-tuning kis masle ke liye sab se munasib hai?​

A) Jawab ka mustaqil andaz ya saakht format chahiye

B) Roz badalte naye facts yaad karwane ke liye

  • C) Internet par search karne ke liye

D) Context window barhane ke liye

Q13. Agent ke run loop me max_turns ka maqsad kya hai?​

A) Jawab ko chhota karna

  • B) Model badalna

C) Loop ko had se zyada chalne se rokna (kharcha aur chakkar se bachat)

  • D) Memory saaf karna
Q14. Handoff aur agent-as-tool me farq kaun sa hai?​

A) Handoff me manager control rakhta hai; as_tool me specialist ko control milta hai

  • B) Handoff me control specialist ko chala jata hai; as_tool me manager control rakhta hai

  • C) Dono bilkul ek jaise hain

D) Handoff sirf tools ke liye hai

Q15. Prompt injection kya hai?​

A) API key ka bohat lamba hona

  • B) GPU ka memory error

  • C) Tokenizer ka ghalat chalna

D) Bahar ke text (web page/email) me chhupi hidayat jo model ko behka de

​
HARD​
Q16. Is code ka output kya hoga?​
import numpy as np
mask = np.tril(np.ones((3, 3), dtype=bool))
print(mask.sum())

A) 6 B) 3

C) 9 D) 0

Q17. Multi-Head Attention me d_model = 32 aur h = 4 heads hon to (bias ke baghair) Wq, Wk, Wv, Wo ke kul parameters kitne?​

A) 1024 B) 16384 C) 4096

D) 512

Q18. Ek 1000 x 1000 weight matrix par LoRA (rank r = 4) ke A aur B ke kul parameters kitne honge? A) 4,000 B) 8,000 C) 1,000,000 D) 4,000,000

Q19. Is code ka output kya hoga?​
def cost(i, o):
return i/1_000_000*1.00 + o/1_000_000*4.00
print(round(cost(2000, 500), 3))

A) 0.002 B) 0.006 C) 0.04 D) 0.004

Q20. Agent ka har qadam 90% sahi hota hai. 5 qadam ke baad poora kaam theek hone ki probability taqreeban kitni?

A) 0.90 B) 0.45 C) 0.59 D) 0.50

Modules 6 aur 9 Answer Key (Jawabat aur wazahat)​

#LevelJawabWazahat
1EasyC) Tokens ko ek sath (parallel)
process kiya aur attention se
rishte pakre
Transformer me recurrence nahi, attention hai: sab tokens ek sath
(parallel) dekhe jate hain, jis se training tez aur door ke rishte aasan hote
hain.
2EasyA) Main kya dhoond raha hoon
(talash)
Query = talash, Key = pehchan (label), Value = asal maal. Query ka Key se
milan weights banata hai.
3EasyD) Agla token sahi andaza
lagana
LLM bohat bara language model hai jo cross-entropy se agla token
andaza lagana seekhta hai; baqi salahiyatein isi se aati hain.
4EasyB) Isay environment variable
ya Colab Secrets me rakhna
chahiye
API key ek secret hai. Chori hone par doosra tumhare paiso se istemal
karega, is liye environment/Secrets istemal karo.
5EasyC) Tumhara apna codeModel sirf function ka naam aur arguments likh kar batata hai; function
tumhara code chalata hai aur natija model ko wapas deta hai.
6ModerateB) Taake softmax saturate na
ho (ek hi weight 1 ke qareeb
na pohanche)
d_k barhne par dot products ka spread barhta hai aur softmax sakht ho
jata hai; sqrt(d_k) se bhaagne par weights narm rehte hain.
7ModerateD) Attention ko tokens ki tartib
khud nahi pata hoti
Self-attention permutation-equivariant hai: tokens shuffle karne se sirf
output ki jagah badalti hai. Isliye position ki maloomat alag se jodte hain.
#LevelJawabWazahat
8ModerateA) BERT encoder-only (dono
taraf dekhta, masked LM); GPT
decoder-only (sirf pichla, agla
token)
BERT bidirectional encoder hai jo text samajhne me acha hai; GPT causal
decoder hai jo agla token andaza lagakar text banata hai.
9ModerateC) Jawab zyada random aur
alag alag ho jate hain
Temperature chhota = probabilities nukeeli (pakka jawab); bara =
barabar (random). Ye zehaanat nahi, sirf randomness badalta hai.
10ModerateD) Taaza/apne documents se
jawab aur sources, jis se
hallucination kam hota hai
RAG pehle sahi chunks dhoondta hai aur prompt me deta hai, is liye
jawab apne data par mabni hota hai aur sources ke saath tasdeeq ho
sakti hai.
11ModerateB) Few-shot promptingFew-shot me misalein prompt ke andar hoti hain aur model ke weights
nahi badalte; fine-tuning me weights badalte hain.
12ModerateA) Jawab ka mustaqil andaz ya
saakht format chahiye
Fine-tuning andaz, format aur behavior seekhati hai. Naye aur badalte
facts ke liye RAG behtar hai.
13ModerateC) Loop ko had se zyada chalne
se rokna (kharcha aur chakkar
se bachat)
max_turns hifazati had hai: agar agent tools ke chakkar me phans jaye to
MaxTurnsExceeded error aata hai aur paise nahi jalte.
14ModerateB) Handoff me control
specialist ko chala jata hai;
as_tool me manager control
rakhta hai
Handoff: naya agent baat sambhalta hai. Agent-as-tool: specialist ek
madadgar kaam karta hai aur manager sab jod kar jawab deta hai.
15ModerateD) Bahar ke text (web
page/email) me chhupi hidayat
jo model ko behka de
Agent jo untrusted text parhta hai us me hamlawar ki hidayat ho sakti
hai. Bahar ke text ko data samjho, hukm nahi, aur tools ki ijazat
mehdood rakho.
16HardA) 6np.tril neeche ka tikon (diagonal samet) True rakhta hai: 1 + 2 + 3 = 6
True values.
17HardC) 4096Parameters 4 x d_model x d_model = 4 x 32 x 32 = 4096. Heads sirf in
matrices ko baant kar istemal karte hain, is liye h se parameters nahi
barhte.
18HardB) 8,000A: (r x d) = 4 x 1000 = 4000, B: (d x r) = 1000 x 4 = 4000. Kul 8000, jab ke
full matrix 1,000,000 parameters ka hai (sirf 0.8%).
19HardD) 0.004Input: 2000/1e6 x 1 = 0.002. Output: 500/1e6 x 4 = 0.002. Jama = 0.004
dollar.
20HardC) 0.59p^n = 0.9^5 = 0.59049, yani sirf taqreeban 59%. Lambe agents ko chhote
qadam, jaanch aur wapsi ka raasta chahiye.

Tumhara score: ______ / 20