IIT Delhi Certificate Programme in Generative AI

• Basic Text Processing (NLTK, spaCy), Morphology, Stemming, Edit Distance

Learning Outcomes:

• Use of NLP tools for text processing

• Language Modelling: N-gram Modelling, Smoothing Techniques, Perplexity

Learning Outcomes:

• Learning different language modelling approaches

• POS Tagging: Sequential Learning, HMM, Viterbi Algorithm

Learning Outcomes:

• Introduction to sequence learning for NLP

• Parsing: Constituency vs Dependency Parsing, CKY Algorithm, CFG, PCFG

Learning Outcomes:

• Learning techniques for syntactic analysis in NLP

• Text Classification: Naive Bayes Algorithm, Lexical Similarity (Word Embeddings, TF-IDF, Word2Vec, GloVe)

Learning Outcomes:

• Understanding classification in NLP using traditional and modern approaches

• Neural Language Models (CNN, RNN, LSTM, GRU, Seq2Seq)

Learning Outcomes:

• Introduction to neural models and attention mechanisms for text generation

• Attention Mechanism: Self-Attention, Transformer Architecture

Learning Outcomes:

• Understanding self-attention and Transformer architectures for language models

• Pre-trained Models: BERT, GPT, T5

Learning Outcomes:

• Architecture and training of different pre-trained large language models

• Fine-tuning Strategies: Task-specific Fine-tuning, Instruction Fine-tuning, Preference Tuning (RLHF, PPO)

Learning Outcomes:

• Learning various fine-tuning approaches to improve model performance

• Prompting Strategies: In-context Learning, Chain of Thought, Knowledge Probing, Text Generation

Learning Outcomes:

• Techniques for optimising LLM performance through effective prompting

• Augmented LLMs: Retrieval-Augmented Generation (RAG), Tool Augmented LLM

Learning Outcomes:

• Leveraging RAG and tool augmentation to improve LLM efficiency and reasoning capabilities

Module 5: Generative AI for Vision

• Vision Language Models (VLM)

Learning Outcomes:

• Understanding how VLMs enable combined text and image generation for multimodal applications

Module 6: Responsible AI

• Misinformation, Bias, Toxicity, Security and Fairness

Learning Outcomes:

• Learning strategies to mitigate bias, toxicity, and hallucinations in AI outputs

• Implement a fully connected neural network using PyTorch or TensorFlow for some task. Train the model and visualise its accuracy and loss over epochs.
• Fine-tune a pre-trained transformer model on a text classification task. Evaluate the model on a test set and report its accuracy, precision, and recall.
• Implement a simplified version of the transformer architecture focusing on self-attention and positional encoding. Train the model on a small dataset for translation or another task. Compare the results with a pre-trained transformer model.

• Implement PEFT on a large LLM for a specific task. Compare the training time and resource usage between standard fine-tuning and PEFT.
• Create a simple reward model using human-labelled data for a text generation task (e.g., summarisation or sentiment correction). Implement a reinforcement learning loop to improve the LLM's responses based on the reward model.
• Compare the performance of different open-source Large Language Models (LLMs) on reasoning tasks using various prompting techniques—zero-shot, few-shot, chain-of-thought, self-consistency prompting.

• Compute matrix operations using NumPy.
• Preprocess a text dataset and generate embeddings using spaCy or Hugging Face.
• Build a simple neural network using TensorFlow/PyTorch.
• Train the model on basic NLP tasks like text classification problem.
• Implement a simplified self-attention mechanism.
• Use a pre-trained transformer models for various NLP tasks like translation or summarization.

• Fine-tune a GPT or BERT model on a custom dataset for various NLP problems
• Use Hugging Face's pre-trained transformer model to generate text.
• Apply PEFT to a large model for fine-tuning on a small dataset.
• Apply quantization and pruning on a pre-trained model to optimize for lower resource usage.
• Implement a RAG model using Hugging Face transformers and document retrieval from a custom corpus.
• Use Hugging Face multilingual models for problems in cross-lingual scenarios.