AI Path Finder
Free consultation

Lesson · Hinglish · ⏱ 7 min padhne mein

📝 Train/test split aur overfitting

EnglishHinglish

🏠 Asli zindagi ki misaal

Jo student pichhle saal ka question paper shabd-shabd ratt leta hai, woh us paper mein 100% laata hai, lekin naye sawaalon wale asli exam mein fail ho jaata hai. Yeh hai overfitting. Teachers asli exam ke sawaal secret rakhkar isse rokte hain. ML mein hum bhi yahi karte hain: ek 'test set' chhupa ke rakhte hain jo model training ke dauraan kabhi nahi dekhta.

💡 Aasaan shabdon mein

Training se pehle data split karo: aam taur pe 70–80% training ke liye aur 20–30% testing ke liye. Model sirf training set se seekhta hai, aur aap use sirf test set pe judge karte ho, jisse pata chalta hai ki woh naye, real-world data pe kaisa karega. Overfitting ka matlab hai training scores bahut acche lekin test scores kharab: model ne pattern seekhne ki jagah noise ratt liya. Underfitting ka matlab dono pe kharab scores: model bahut simple hai. Cross-validation split ko kai baar repeat karta hai, taaki estimate zyaada reliable ho.

🎬 Dekho kaise kaam karta hai

Train · 80% · the model studies theseTest · 20% 🔒Train score 0.99Test score 0.71Big gap → overfitting (memorised, not learned)

Pehle data split karo: model 80% se seekhta hai aur us 20% pe judge hota hai jo usne kabhi nahi dekha. Dono scores mein bada gap matlab overfitting.

🪜 Step by step

  1. Sabse pehle split karo: lagbhag 70% train, 15% validation, 15% test.
  2. Sirf training set pe train karo, aur cleaning steps (jaise scaling) bhi usi pe fit karo.
  3. Models aur settings ko validation set pe compare karo.
  4. Fark dekho: agar training validation se kaafi behtar hai, to overfitting hai. Simple karo ya data badhao.
  5. Test set ek hi baar use karo, aakhir mein, jo number report karna hai uske liye.

🔑 Zaroori shabd

Test set
Training ke dauraan chhupa hua data, jo sirf final check ke liye use hota hai.
Overfitting
Training data ratt lena aur naye data pe fail hona.
Cross-validation
Train/test split kai baar repeat karna, taaki score pe zyaada bharosa ho.

👀 Example dekho

python
from sklearn.model_selection import train_test_split
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2, random_state=42)

model.fit(X_train, y_train)
print("train:", model.score(X_train, y_train))   # 0.99
print("test: ", model.score(X_test, y_test))     # 0.71  ← overfitting alarm 🚨

Train aur test scores ke beech bada gap overfitting ka classic signal hai.

⚠️ Aam galtiyaan jinse bachna hai

  • Time based data ke liye random split. Past pe train karo, future pe test, jaise asli use mein hota hai.
  • Ek hi customer train aur test dono mein. Model customer ko yaad kar leta hai; customer ke hisaab se split karo.
  • Tuning karte hue baar baar test score dekhna. Phir woh imaandaar test nahi rehta.

✅ Jaldi se check karo

Training accuracy 98% hai aur test accuracy 65%. Kya ho raha hai?

  1. Model perfect hai
  2. Underfitting
  3. Overfitting
Jawaab dekho

C. Overfitting: Dekhe hue data pe bahut accha aur naye data pe kharab, matlab model ne seekha nahi, ratt liya.

🙋 Seekhne wale yeh poochhte hain

Cross validation kya hai?

Training data ko k hisson mein baantna aur baari baari ek hissa validation ke liye rakhna. Data kam ho to zyaada stable estimate deta hai.

Underfitting kaise theek karun?

Model bahut simple hai: behtar features daalo, zyaada powerful model lo, ya zyaada der train karo.

📌 Yaad rakho

Model ko kabhi us data pe judge mat karo jis pe woh train hua hai.

📖 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 · ML Engineer · MLOps / AI Platform Engineer · AI Research Engineer

Course mein seekho