Custom LLM Development: Fine-Tuning, QLoRA, RAG & AI Application Deployment
Build Domain-Specific Large Language Models Using Python, Hugging Face, Efficient Fine-Tuning & Retrieval-Augmented Generation
Skills you will gain:
About Workshop:
Move beyond using ready-made AI tools and learn how to customize, fine-tune, evaluate and deploy open-source Large Language Models (LLMs). This advanced, practical workshop introduces the complete LLM customization workflow in a focused three-day format: selecting a suitable model, preparing domain-specific datasets, applying parameter-efficient fine-tuning, evaluating and optimizing model outputs, and deploying a working AI application. Participants will work with modern open-source models and widely used development frameworks.
Aim:
Designed for the next generation of AI innovators, this workshop aims to take participants beyond conventional prompt engineering into the world of Advanced LLM Engineering and Agentic AI Development. It focuses on the practical transformation of open-source Large Language Models into customized, high-performance AI applications using LoRA, QLoRA, custom data fine-tuning, RAG pipelines, model optimization and intelligent AI agents.
Workshop Objectives:
- Understand advanced LLM architectures and Generative AI workflows.
- Prepare custom datasets for LLM fine-tuning.
- Apply LoRA, QLoRA and PEFT techniques.
- Explore model alignment, evaluation and optimization.
- Develop Retrieval-Augmented Generation (RAG) applications.
- Build intelligent AI agents using tool calling.
- Gain practical exposure to LLM deployment and API integration.
What you will learn?
Day 1: LLM Foundations, Model Selection & Custom Dataset Engineering
- Explore modern open-source LLMs, including Llama, Mistral, Qwen, and Phi, and their real-world applications.
- Understand Transformer architecture, tokenization, context windows, and base versus instruction-tuned models.
- Compare full fine-tuning, LoRA, QLoRA, and DoRA for efficient model customization.
- Select suitable LLMs based on model size, hardware requirements, licensing, and domain-specific tasks.
- Prepare custom instruction datasets using JSONL, conversational formats, and model-specific chat templates.
- Apply dataset cleaning, deduplication, quality validation, and train/validation/test splitting to prevent data leakage.
Hands-On Activity: Load and explore a lightweight open-weight LLM using Hugging Face and Google Colab. Prepare, clean, and validate a domain-specific instruction dataset and develop a Custom LLM Training Dataset Pipeline.
Tools: Python | Google Colab | Hugging Face Transformers | Hugging Face Datasets | Pandas | PyTorch | JSONL
Day 2: Hands-On LLM Fine-Tuning, QLoRA Optimization & Model Evaluation
- Understand the end-to-end LLM fine-tuning workflow using Hugging Face Transformers, PEFT, TRL, and Unsloth.
- Configure LoRA adapters, including rank, alpha, target modules, and parameter-efficient training strategies.
- Apply QLoRA and 4-bit quantization for memory-efficient LLM fine-tuning.
- Perform Supervised Fine-Tuning (SFT) using custom domain-specific instruction datasets.
- Configure learning rate, batch size, epochs, gradient accumulation, and model checkpoints.
- Monitor training loss and evaluate baseline versus fine-tuned model outputs using held-out test prompts.
- Explore hallucination detection, response-quality assessment, and introductory preference optimization techniques, including DPO.
Hands-On Activity: Fine-tune a selected open-weight LLM using QLoRA, PEFT, and Unsloth in Google Colab. Monitor training performance, save adapter weights, compare model responses, and develop a Fine-Tuned LLM Performance Evaluation Report.
Tools: Google Colab | Python | Hugging Face Transformers | PEFT | TRL | Unsloth | bitsandbytes | PyTorch
Day 3: LLM Optimization, RAG Integration & AI Application Deployment
- Explore fine-tuned model inference, adapter loading, model merging, and quantization for efficient deployment.
- Understand local and scalable LLM serving using Ollama and vLLM.
- Develop LLM-powered API endpoints using Python and FastAPI.
- Explore Retrieval-Augmented Generation (RAG), embeddings, semantic search, and vector databases.
- Build document-based retrieval workflows using FAISS and integrate retrieved context with LLM responses.
- Understand structured outputs, function calling, and introductory Agentic AI workflows.
- Examine LLMOps fundamentals, including model versioning, monitoring, evaluation, data privacy, and deployment security.
Hands-On Activity: Load a compatible fine-tuned or preconfigured model, build a RAG-enabled Domain-Specific AI Assistant using FAISS, connect it through FastAPI, and evaluate responses using sample documents and domain-specific queries.
Tools: Python | Hugging Face | FastAPI | FAISS | Google Colab | Ollama (Demo) | vLLM (Demo) | LangChain (Optional)
Mentor Profile
Fee Plan
Important Dates
19 Nov 2026 Indian Standard Timing 4:30 PM IST
17 Nov 2026 to 19 Nov 2026 Indian Standard Timing 5:30PM
Get an e-Certificate of Participation!
Intended For :
Career Supporting Skills
Workshop Outcomes
- Understand modern open-source LLM architectures and model-selection considerations.
- Choose between prompting, RAG and fine-tuning for different application needs.
- Prepare and validate custom instruction datasets for model adaptation.
- Apply LoRA or QLoRA in a guided supervised fine-tuning workflow.
- Understand preference optimization concepts such as DPO.
- Evaluate base and fine-tuned models using appropriate quality measures.
- Understand quantization and inference optimization options.
- Build a basic API and explore RAG or tool integration for a customized LLM.