nndl/llm-beginner

Features

• Agentic LLM Frameworks - Provides a comprehensive framework for building autonomous agents featuring reasoning loops, tool use, and memory management.
• Reasoning-Action Loops - Implements a cycle of thought, action, and observation for tool use and autonomous self-correction.
• Agentic Reasoning Loops - Implements iterative reasoning and execution cycles that persist state until task completion.
• Coding Agents - Builds agent systems that automate code modification and testing within local repositories.
• Development Environment AI Integrations - Integrates models with development environments to modify source code and verify fixes via test suites.
• Subagent Coordination - Deploys specialized child agents with isolated contexts to handle parallel sub-tasks efficiently.
• Autonomous Coding Agents - Develops an autonomous agent that reads issues and modifies code iteratively to resolve software bugs.
• Autonomous Agent Loops - Creates autonomous loops that enable models to use external tools for multi-step task completion.
• Preference Alignment - Refines model outputs through preference alignment using pairs of chosen and rejected responses.
• Direct Preference Optimization - Optimizes models using direct preference optimization to align outputs with chosen and rejected response pairs.
• Decoder Architectures - Implements decoder-only transformer architectures featuring rotary positional embeddings and causal attention.
• Instruction Fine-tuning - Implements supervised instruction fine-tuning using low-rank adaptation to create chat-capable assistants.
• KV Cache Management - Accelerates autoregressive generation by caching previously computed key and value tensors.
• RAG Pipelines - Builds retrieval pipelines using embedding models and vector stores to answer questions from external documents.
• Alignment Techniques - Guides the adaptation of base models into assistants using supervised tuning, LoRA, and preference optimization.
• Low-Rank Adaptation - Provides parameter-efficient fine-tuning by injecting trainable low-rank matrices into linear layers.
• Language Model Training - Implements language model training using next-token prediction and learning rate scheduling to minimize perplexity.
• Causal Masking - Implements upper-triangular attention masking to prevent models from attending to future tokens.
• Rotary Positional Embeddings - Encodes token positions using rotation matrices to maintain relative distance across sequence lengths.
• RAG Implementations - Constructs end-to-end systems that extract text, index it via embeddings, and generate grounded answers.
• Retrieval Augmented Generation Pipelines - Constructs end-to-end pipelines that combine vector search and document indexing to provide grounded answers.
• Supervised Fine-Tuning - Trains models on conversational datasets using next-token prediction and loss masking.
• Transformer Architecture Implementation - Provides technical instructions for building generative and encoder transformer models from scratch.
• Transformer Blocks - Implements self-attention mechanisms and encoder blocks for text classification and language modeling.
• Transformer Encoders - Implements transformer encoder architectures from scratch to perform text classification tasks.
• Vector Similarity Search - Implements a full pipeline for retrieving relevant external context using high-dimensional vector embeddings and similarity search.
• ReAct Agents - Implements the ReAct paradigm, allowing models to iteratively reason and act using external tools.
• Decoder-Only Architectures - Guides the construction of generative models featuring tokenization, rotary embeddings, and KV-caching for autoregressive generation.
• Model Fine-Tuning - Transforms base models into chat assistants using supervised fine-tuning and preference optimization.
• LLM Education - Provides a comprehensive educational curriculum covering LLM architectures, training methods, and deployment strategies.
• Transformer Implementations - Provides pedagogical implementations of transformer models from scratch, including self-attention and encoder blocks.
• Autonomous Coding Agents - Develops autonomous agents capable of reading repositories, modifying source code, and verifying fixes through test execution.
• Agentic Error Recovery - Provides mechanisms to feed tool exceptions back into the model for autonomous self-correction.
• Automated Skill Loading Systems - Provides a system for dynamically discovering and loading skill definitions from markdown files into an agent's context.
• RAG Document Retrieval - Performs top-k retrieval using query embeddings and refines results with a reranker.
• External Tool Integrations - Connects models to specialized functions and APIs using structured schemas to extend capabilities.
• Grounded Answer Generation - Combines retrieved document segments into prompts to ensure responses are grounded in provided context.
• KV Cache Optimizations - Optimizes key-value caches to increase sampling speed during autoregressive generation.
• Faithfulness Verifiers - Verifies that generated answers are strictly supported by retrieved sources to prevent hallucinations.
• Model Comparison Interfaces - Provides a framework for side-by-side comparison of outputs between base and fine-tuned model versions.
• Model Evaluation Metrics - Implements methods for measuring model performance through retrieval metrics, faithfulness checks, and attention visualization.
• Model Context Protocol Servers - Implements Model Context Protocol servers to expose file manipulation and version control tools.
• Recursive Text Splitting - Splits raw text into segments using recursive or semantic methods for vector search preparation.
• RAG Evaluation Frameworks - Evaluates RAG system quality using automated judges to measure recall and faithfulness.
• RAG Tutorials - Provides a step-by-step guide to building RAG pipelines using vector stores, embeddings, and reranking.
• Cross-Encoder Rerankers - Refines retrieval precision by scoring document-query pairs using a cross-encoder architecture.
• Text Model Training - Trains transformer-based models on sentiment datasets for text classification tasks.
• Vector Retrieval Systems - Creates searchable indices using embeddings and vector stores to retrieve relevant document chunks.
• AI Observability and Evaluation - Provides observability and evaluation tools for monitoring agent performance through structured traces.
• Vector Indexing - Generates document embeddings and stores them in high-dimensional indexes for semantic search.
• Custom Tool Definitions - Provides a framework for defining modular, typed tool schemas and execution logic for agents.
• Agent Implementation Guides - Offers a practical guide for developing ReAct loops, tool integration, and autonomous agentic workflows.
• Fine-Tuning Guides - Serves as a technical resource for implementing supervised fine-tuning and preference optimization for model alignment.
• Transformer Tutorials - Provides a technical walkthrough for implementing encoder and decoder transformer architectures from scratch.
• Goal Accuracy Evaluators - Measures agent success rates using keyword validation and trace analysis to verify goal achievement.
• Retrieval Metrics - Calculates retrieval quality using gold standard datasets to determine recall and mean reciprocal rank.

Star history