Gen AI & LLM for Data Scientists

Describe the basic architecture of a Recurrent Neural Network (RNN) for processing sequential data.

Explain how LSTM and GRU cells use gating mechanisms to improve upon simple RNNs.

Identify the limitations of RNNs, such as the vanishing gradient problem and difficulty with long-range dependencies.

Describe the attention mechanism as a weighted sum of values based on query-key similarity.

Compute attention scores given query, key, and value matrices to build intuition.

Explain byte-pair encoding (BPE) tokenization and its role in handling large vocabularies in LLMs.

Explain the role and implementation of positional encoding in transformers.

Describe scaled dot-product self-attention and articulate why scaling is necessary.

Explain multi-head attention and its advantages over single-head attention.

Describe the role of feed-forward networks and layer normalization in a transformer block.

Trace the data flow through a transformer encoder stack.

Explain the role of causal masking in the decoder for autoregressive generation.

Trace the data flow through a full encoder-decoder transformer model.

Explain the concept of transfer learning in NLP and why pre-training is effective.

Describe BERT's masked language modeling (MLM) and next sentence prediction (NSP) training objectives.

Describe GPT's autoregressive causal language modeling objective.

Compare encoder-only, decoder-only, and encoder-decoder architectures and their typical use cases.

Fine-tune a pre-trained BERT model on a text classification task using the Hugging Face library.

Explain scaling laws relating model size, data, and compute to performance.

Describe the pre-training phase of an LLM, including its data requirements and compute scale.

Describe emergent abilities observed in large language models that are not present in smaller models.

Compare the key characteristics of major open-source and closed-source LLM families (e.g., GPT, Llama, Mistral, Claude).

Explain the purpose of the alignment phase in LLM training.

Describe supervised fine-tuning (SFT) or instruction tuning for teaching models to follow instructions.

Explain the three main steps of Reinforcement Learning from Human Feedback (RLHF): reward model training, fine-tuning, and PPO.

Compare RLHF with newer alignment techniques like Direct Preference Optimization (DPO).

Describe safety techniques used in alignment, such as red-teaming and Constitutional AI.

Describe and contrast decoding strategies: greedy, temperature sampling, top-k, and nucleus sampling.

Interact with an LLM via an API (e.g., OpenAI, Hugging Face) to perform generation tasks.

Apply zero-shot and few-shot prompting to solve a classification or Q&A task.

Apply chain-of-thought prompting to improve reasoning in LLM outputs.

Structure system prompts and user/assistant roles to elicit consistent behavior from an LLM.

Evaluate LLM outputs using both manual frameworks (fluency, relevance) and automated metrics (ROUGE, BLEU).

Explain the limitations of averaging word embeddings for sentence representation.

Describe the Sentence-BERT (SBERT) architecture and its siamese network training approach for creating meaningful sentence embeddings.

Encode a collection of documents into embeddings using a sentence transformer model.

Explain the need for Approximate Nearest Neighbor (ANN) search for efficient retrieval at scale.

Set up a local vector store (e.g., FAISS or ChromaDB) and index document embeddings.

Implement a function to encode a text query and retrieve the top-k most similar documents from a vector store.

Explain how RAG addresses LLM limitations such as hallucination and knowledge cutoffs.

Describe the end-to-end RAG architecture: indexing, retrieval, and generation.

Apply and contrast different document chunking strategies (e.g., fixed-size, recursive, semantic).

Implement a re-ranking step to improve the relevance of retrieved documents.

Integrate retrieved context into an LLM prompt to generate a grounded answer.

Evaluate a RAG pipeline using metrics such as faithfulness, answer relevance, and context relevance.

Distinguish between scenarios that call for prompt engineering versus those requiring model fine-tuning.

Explain the concept of parameter-efficient fine-tuning (PEFT) and its benefits over full fine-tuning.

Describe how Low-Rank Adaptation (LoRA) modifies model weights efficiently during training.

Prepare an instruction-following dataset for a fine-tuning task.

Apply a LoRA configuration to a pre-trained model using a library like Hugging Face's PEFT.

Execute a fine-tuning job and evaluate the adapted model's performance against a baseline.

Explain the concept of function calling (tool use) in LLMs.

Describe the ReAct (Reasoning + Acting) framework for enabling agentic LLM behavior.

Implement a basic agent that uses a predefined tool to answer a question it cannot answer on its own.

Describe the core abstractions and design philosophies of LangChain and LlamaIndex.

Build a simple RAG application using a framework like LangChain or LlamaIndex.

Describe the concept of multi-agent architectures and their potential use cases.

Explain recent architectural innovations in LLMs, such as Mixture of Experts (MoE) and Grouped-Query Attention (GQA).

Describe the high-level mechanism of diffusion models for image generation (e.g., Stable Diffusion).

Explain how multimodal models like GPT-4V and LLaVA process both text and image inputs.

Summarize the current capabilities and key models in generative audio (e.g., speech-to-text, text-to-speech).

Identify key trends and breakthroughs in generative AI from the past 12-18 months.

Identify reliable sources (e.g., key conferences, blogs, researchers) for staying updated on AI developments.

Discuss key ethical considerations in the development and deployment of generative AI systems.