Mixed Self-Check

Answer these without looking back

Try each question aloud, then check the section in brackets. (1) A tokenizer gives [CLS] play ##ing [SEP]; which algorithm family and which model family does this suggest? (Tokenizers.) (2) Why must attention_mask accompany padded batches? (Tokenizers.) (3) What warning tells you a classification head was randomly initialised, and when is it expected? (Models.) (4) Why slice out[0, prompt_len:] before decoding? (Generation.) (5) When would you choose beam search over sampling? (Generation.) (6) What does batched=True change in Dataset.map? (Datasets.) (7) Where does Trainer get the loss? (Fine-tuning.) (8) Why is accuracy misleading on imbalanced classes? (Metrics.) (9) How many bytes per parameter for bfloat16 and for 4-bit weights? (Quantization.) (10) Name two reasons to prefer safetensors and a pinned revision. (Security.) If any answer felt shaky, rerun that section's code and quiz before moving on.

A last end-to-end exercise

Fill in the blanks from memory, then compare with the earlier sections.

from transformers import pipeline

# 1. Pick a task and a model from the Hub (pin a revision in real projects)
gen = pipeline("text-generation", model="Qwen/Qwen2.5-0.5B-Instruct")

# 2. Format the conversation as messages, not raw strings
messages = [{"role": "user", "content": "Summarise LoRA in one sentence."}]

# 3. Control length and randomness explicitly
reply = gen(messages, max_new_tokens=60, do_sample=True, temperature=0.7)

# 4. Read the last message: the assistant's reply
print(reply[0]["generated_text"][-1]["content"])

Spaced revision works

Come back to this self-check after one day, then after a week. Questions you can answer instantly the third time are truly learned; the rest point to the sections worth a second read.

Quick check: You must run a 7B chat model on one 16 GB GPU and adapt it to your college's FAQ data. Which plan is most realistic?

• Load it in 4-bit and train LoRA adapters (QLoRA), then evaluate on held-out FAQs
• Full fine-tuning in float32
• Train a 7B model from scratch on the FAQ data
• Use zero-shot classification on the FAQ answers