Generative AI & LLMs exam prep

Exam

Fri 11 Dec 2026 54 days to go

Registration

Closes 31 Oct 2026 ₹750 fee · exam date may change, check the hall ticket

Score

25% assignments + 75% exam Best 3 of 7 assignments count

To pass

Exam ≥ 30/75 Plus assignments ≥ 10/25 · Gold ≥ 90, Silver 76–89, Bronze 61–75

Where the marks are. Weeks 3, 5 and 6 carry the most testable detail: scaled dot-product attention, decoding settings (temperature, top-k, top-p), prompting styles, LoRA and other PEFT methods, the RLHF pipeline, RAG and evaluation metrics. Learn the formulas on the Numericals page; they turn into easy 2-mark questions.

• Stemming chops suffixes (may not give a real word); lemmatization returns the dictionary form.
• A word in every document has IDF = 0, so TF-IDF = 0.
• One-hot vectors carry no similarity; all pairs are orthogonal.
• CBOW: context → word. Skip-gram: word → context (better for rare words).
• Word2Vec/GloVe are static; BERT embeddings are contextual.
• Vanishing gradients → LSTM/GRU; exploding gradients → gradient clipping.
• LSTM has 3 gates + a cell state; GRU has 2 gates and no separate cell.
• Teacher forcing feeds the true previous token during training.
• Beam width 1 = greedy.
• Scale by √d_k, not d_k. Attention weights sum to 1.
• Transformers need positional encoding; attention alone is order-blind.
• Decoder self-attention is masked; cross-attention takes Q from the decoder, K/V from the encoder.
• Self-attention is O(n²) in sequence length but runs in parallel.
• BERT = encoder, masked LM, bidirectional. GPT = decoder, next token. T5 = encoder–decoder.
• Discriminative learns P(y|x); generative learns P(x) / P(x,y).
• VAE loss = reconstruction + KL; the reparameterisation trick makes sampling differentiable.
• GAN failure with little variety = mode collapse.
• Diffusion: forward adds noise (fixed), reverse removes it (learned).
• CLIP is contrastive image–text training, enabling zero-shot classification.
• Few-shot prompting changes no weights.
• Temperature up = more random; T → 0 = greedy.
• Top-p picks a variable number of tokens; top-k a fixed number.
• Full fine-tuning risks catastrophic forgetting; PEFT reduces it.
• LoRA trainable params = r(d + k); base weights are frozen; no added inference latency once merged.
• Prompt tuning trains soft prompt embeddings, not the words you type.
• RLHF order: SFT → reward model → PPO. The KL penalty prevents reward hacking.
• DPO needs no separate reward model.
• ROUGE → summarization (recall); BLEU → translation (precision).
• Lower perplexity is better.
• RAG adds knowledge without retraining; fine-tuning changes behaviour/style.
• LangChain chain = fixed steps; agent = LLM chooses tools.
• Quantisation = lower precision; distillation = student learns from teacher; pruning = remove weights.
• Chinchilla: about 20 tokens per parameter.
• Hallucination is reduced, never eliminated, by RAG and low temperature.
• Red-teaming = deliberately attacking the model to find failures.
• EU AI Act is risk-based: unacceptable, high, limited, minimal.

• Word2Vec — Mikolov et al., 2013CBOW and skip-gram.
• Neural Machine Translation by Jointly Learning to Align and Translate — Bahdanau et al., 2014The first attention for seq2seq.
• Attention Is All You Need — Vaswani et al., 2017The transformer.
• BERT — Devlin et al., 2018Masked language modelling, encoder-only.
• Language Models are Few-Shot Learners (GPT-3) — Brown et al., 2020In-context learning at 175B parameters.
• Generative Adversarial Networks — Goodfellow et al., 2014GANs.
• Auto-Encoding Variational Bayes — Kingma & Welling, 2013VAEs.
• Denoising Diffusion Probabilistic Models — Ho et al., 2020Diffusion. Latent diffusion: Rombach et al., 2022.
• CLIP — Radford et al., 2021Contrastive image–text pre-training.
• LoRA — Hu et al., 2021Low-rank adaptation.
• InstructGPT — Ouyang et al., 2022The RLHF recipe.
• Chinchilla — Hoffmann et al., 2022Compute-optimal training.
• Retrieval-Augmented Generation — Lewis et al., 2020RAG.