Large Language Models

This hub is designed to bridge the gap between theoretical concepts and practical implementation of LLMs. Whether you’re a researcher, developer, or enthusiast, you’ll find structured pathways to master cutting-edge techniques like RAG, fine-tuning, and neuro-symbolic AI, all demonstrated through offline, reproducible code using open-source models like Llama-3.

1. Core LLM Concepts

Foundational Knowledge for Building and Customizing LLMs

1.1 Self-Attention & Transformers

Why It Matters: Self-attention is the backbone of transformer models, enabling LLMs to process context and relationships in text.

Revise Self-Attention Mechanisms
- Content:
  - Step-by-step visualization of how input text is split into Query, Key, and Value vectors.
  - How these vectors interact to produce a context-aware output.
  - Practical analogies to simplify mathematical operations (e.g., dot products, softmax).
- Learning Outcome: Understand how transformers capture long-range dependencies in text.
- Video Tutorial : https://youtu.be/L_bBglaRPfo?si=RxM3Q48UyIGiI_-A
  - Revise Sparse-Attention and Cross-Attention Mechanisms
    - Content:
      - Step-by-step visualization of how it is different from Vanilla Self Attention.
      - How these vectors interact to produce a context-aware output.
      - Practical analogies to simplify mathematical operations.
    - Learning Outcome: Understand how to utilize these attention mechanisms.
    - Video Tutorial Sparse-Attention: https://youtu.be/fWto5Ozpjsc?si=0ius7ETUO2uQvO0k
    - Video Tutorial Cross-Attention: https://youtu.be/WfJ8waoakeQ?si=iH9KK8hc-9ZI34TN
Transformers for Autoregressive Models
- Content:
  - Role of the decoder architecture in models like GPT.
  - How masked self-attention enables autoregressive text generation (predicting the next token).
- Learning Outcome: Connect transformer mechanics to real-world applications like chatbots.
- Video Deep Dive: https://youtu.be/KNoW9E-TDU8?si=l-iOYy2z7tekEEFZ

1.2 Handling Long Text Sequences

Why It Matters: Most LLMs struggle with long inputs. Learn modern solutions to this limitation.

Part 1: Vanilla Transformer Limitations
- Content:
  - Why positional encoding fails for sequences longer than the training context.
  - The "attention collapse" problem in long texts.
- Video : https://www.youtube.com/watch?v=q2otBk4Wcx8
Part 2: Attention with Linear Biases (ALiBi)
- Content:
  - How ALiBi’s linear slope adjusts attention scores for extrapolation.
  - Comparison with traditional positional embeddings.
- Video: https://youtu.be/I04hB_QAjFU?si=rWQ1LQKn-9p8Dzw9
Part 3: Rotary Positional Embeddings (RoFormer)
- Content:
  - Rotary matrices to encode positions without increasing parameters.
  - Why RoFormer outperforms ALiBi in certain tasks.
- Video: https://youtu.be/5WhQecvWX7U?si=8w8d7yYzyFJGMEjB

2. Retrieval-Augmented Generation (RAG)

Enhance LLMs with External Knowledge Bases

2.1 RAG Fundamentals

Why It Matters: RAG combines LLMs with retrieval systems to reduce hallucinations and improve factual accuracy.

Introduction to RAG
- Content:
  - Three-Step Workflow: Retrieve → Augment → Generate.
  - Demo: Build a RAG pipeline using Llama-3 and FAISS for vector search.
  - Code walkthrough for document chunking, embedding, and query augmentation.
- Learning Outcome: Implement a basic RAG system from scratch.
- Video : https://youtu.be/DBprEyQBeKQ?si=HeILh01l6SxkWgBs
- Code Walkthrough: https://www.quantacosmos.com/2024/06/rag-retrieval-augmented-generation-llm.html

2.2 Advanced RAG Techniques

Why It Matters: Basic RAG struggles with complex queries. These methods add structure to retrieval.

Graph-Based RAG (GraphRAG)
- Part 1: Theory
  - Content:
    - Represent documents as knowledge graphs (entities + relationships).
    - Use graph traversal for context-aware retrieval.
  - Video: https://youtu.be/cMXyNGObcGw?si=aLZoME_kQR3FNaj1
  - Code Walkthrough: https://www.quantacosmos.com/2024/06/rag-retrieval-augmented-generation-llm.html
- Part 2: Implementation
  - Content:
    - Offline demo with Llama-3 and NetworkX for graph operations.
    - Querying subgraphs for precise context extraction.
  - Video: https://youtu.be/pbhRFZwmOvU?si=7lXozQwyxkacZ4We
  - Code Walkthrough: https://www.quantacosmos.com/2024/06/rag-retrieval-augmented-generation-llm.html
Knowledge Hypergraphs
- Content:
  - Extend graphs to n-ary relationships (e.g., "Company A acquires Company B for $X in Year Y").
  - Demo: Storing hyperedges in a graph database (e.g., Neo4j).
- Video : https://youtu.be/SPt5O3rpHIo?si=VZuPc_y_Pfs5K0_o
- Code Walkthrough: https://www.quantacosmos.com/2024/06/knowledge-hyper-graph-with-llm-rag.html
Zero-Shot & One-Shot RAG
- Zero-Shot:
  - Content: Answer queries without task-specific training (e.g., "Explain quantum physics to a 5-year-old").
  - Video: https://youtu.be/uEQlGEGKmFU?si=gmypCfw20rT8OOKC
  - Code Walkthrough: https://www.quantacosmos.com/2024/06/zero-shot-llm-rag-with-knowledge-graph.html
- One-Shot:
  - Content: Adapt to custom tasks with a single example (e.g., "Generate a sales email using this template").
  - Video: https://youtu.be/AusPKVSkvGI?si=OICT124ec2_LRUT8
  - Code Walkthrough: https://www.quantacosmos.com/2024/06/one-shot-llm-rag-with-knowledge-graph.html

3. Fine-Tuning & Adaptation

Customize LLMs for Domain-Specific Tasks

3.1 Parameter-Efficient Fine-Tuning (PEFT)

Why It Matters: Full fine-tuning is resource-heavy. PEFT methods reduce costs while retaining performance.

LoRA (Low-Rank Adaptation)
- Content:
  - Inject low-rank matrices into transformer layers.
  - Mathematical intuition behind rank reduction (SVD analogy).
- Video: https://youtu.be/KPJMTN0Lv0M?si=ebOOX-qGrSy9MWkf
- Code Walkthrough: https://www.quantacosmos.com/2024/06/lora-qlora-and-fine-tuning-large.html
QLoRA (Quantized LoRA)
- Content:
  - 4-bit quantization + LoRA for memory-efficient training.
  - Benchmark comparisons: QLoRA vs. LoRA vs. full fine-tuning.
- Video: https://youtu.be/24Px6Gr5uiQ?si=VCdldpU84genKJUo
- Code Walkthrough: https://www.quantacosmos.com/2024/06/lora-qlora-and-fine-tuning-large.html
DORA (Dynamic Low-Rank Adaptation)
- Content:
  - Automatically adjust the rank of LoRA matrices during training.
  - When to prefer DORA over static LoRA.
- Video: https://youtu.be/PAalu1hKTy4?si=QOr_c1MeR8SHRygA
- Code Walkthrough: https://www.quantacosmos.com/2024/07/finetune-large-language-models-with.html

3.2 Full Fine-Tuning Workflows

For High-Resource Scenarios

Fine-Tuning Llama-3 Locally
- Content:
  - Hardware Setup: GPU/CPU requirements, RAM optimization.
  - Data preparation: Formatting instruction datasets (e.g., Alpaca-style).
  - Code: Training loops, checkpointing, and evaluation.
- Video: https://www.youtube.com/watch?v=H1x7Y-6B6Y0
- Code Walkthrough: https://www.quantacosmos.com/2024/06/fine-tune-pretrained-large-language.html

4. Advanced Applications

Innovate with Hybrid AI Systems

4.1 Neuro-Symbolic AI with LLMs

Why It Matters: Combine neural networks’ pattern recognition with symbolic logic’s reasoning.

Algorithmic Trading Case Study
- Content:
  - Symbolic Component: Rule-based market indicators (e.g., moving averages).
  - Neural Component: LLM analyzing news sentiment.
  - Fusion: Decision engine balancing both inputs.
- Video: https://youtu.be/5qEXCxsV4Og?si=3tenzF8wDtcZQohE
- Code Walkthrough: https://www.quantacosmos.com/2025/02/enhancing-algorithmic-trading-with.html

4.2 Quantization for Efficiency

Why It Matters: Deploy LLMs on edge devices (e.g., laptops, phones).

Quantization Basics
- Content:
  - 8-bit vs. 4-bit precision tradeoffs.
  - Tools: GGUF, bitsandbytes, and llama.cpp.
- Video: https://youtu.be/yNNNfFiuKAI?si=9fBEj3EXIRw2_52a

5. Tools & Implementation Guides

Hands-On Support for Real-World Projects

5.1 Local Llama-3 Deployment

Why It Matters: Avoid cloud costs and privacy risks by running models offline.

Step-by-Step Setup
- Content:
  - Downloading Llama-3 weights (via Hugging Face or direct links).
  - Using llama-cpp-python for CPU inference.
  - Optimizing inference speed with Metal (Mac) or CUDA (NVIDIA).
- Video Guide: https://youtu.be/AaoxeuQD-Sg?si=ijxRbynG2B98nvt3