1. Why is PyTorch great for rapid prototyping?

❌ It has intuitive Python syntax
❌ It offers tools for many ML models
❌ It supports diverse neural architectures
✅ It allows real-time testing of code segments without waiting for full implementation

Explanation:
PyTorch’s eager execution means operations run immediately, allowing fast debugging and incremental testing.

---

2. Best fine-tuning method for limited data & low compute?

❌ Train for many epochs on tiny data
❌ Full fine-tuning on irrelevant dataset
✅ Parameter-efficient fine-tuning (PEFT)
❌ Use pretrained model as-is

Explanation:
PEFT (e.g., LoRA, adapters) updates only small portions of the model, preserving general knowledge and reducing compute.

---

3. Essential step when adapting a pretrained model (4 classes → 2 classes)?

❌ Use model without modifying output layer
❌ Freeze all layers except embeddings
❌ Replace transformer with LSTM
✅ Adjust the final output layer from 4 to 2 neurons

Explanation:
The classifier head must match the number of target labels for the new task.

---

4. What does num_labels control in BERT sequence classification?

❌ Controls batch size
❌ Loads tokenizer
✅ Defines output categories & sets number of output neurons
❌ Controls learning rate

Explanation:
num_labels directly determines the size of the classification head (e.g., 3-class, 5-class, etc.).

---

5. Benefit of Hugging Face SFTTrainer?

❌ Tunes hyperparameters automatically
✅ Simplifies training by automating common tasks and reducing errors
❌ Increases dataset size
❌ Creates new tokenization schemes

Explanation:
SFTTrainer handles tokenization, batching, training loops, logging, evaluation, and more—reducing boilerplate PyTorch code.

---

🧾 Summary Table

---