Representation

• Word embeddings, contextual embeddings and what each fails to capture
• BERT-family encoders for classification, tagging and extraction
• Fine-tuning practice: learning rates, epochs and catastrophic forgetting

Tasks

• Named entity recognition, classification and span extraction
• Question answering and summarisation, extractive against abstractive
• Evaluation: F1, ROUGE, BLEU and why each correlates poorly with quality

LLM Systems

• Retrieval-augmented generation and when a smaller fine-tuned model is better
• Structured output, function calling and schema validation
• Cost, latency and the case for not using an LLM at all

Certification

• Serving, batching and monitoring an NLP service in production
• Bias, toxicity and privacy in text data, including PII handling
• A defensible end-to-end project with evaluation evidence