Unsloth is a full-stack, local-first AI platform for fine-tuning and inference of large language models (LLMs), diffusion models, and audio models. It consists of two primary components: (1) the core `unsloth` Python library, which provides highly optimized LoRA/QLoRA/RL training with custom Triton kernels, and (2) the Unsloth Studio application, a Tauri desktop app with a React frontend (TanStack Router, Zustand) and a FastAPI backend (SQLite, structlog). The platform enables users to discover models from the Hugging Face Hub, download and manage them locally, run inference with chat, image/video generation, and RAG, fine-tune with a visual recipe editor, and export in multiple formats (GGUF, NVFP4). It also integrates external LLM providers (OpenAI, Anthropic, Ollama) and supports agentic tool loops via MCP and Codex transports. The codebase is extensively tested with both unit tests and source-level contract tests, and includes cross-platform desktop installers for Windows, macOS, and Linux.
Ray is a distributed compute framework that unifies task/actor execution, data processing, model training, hyperparameter tuning, RL, and model serving under a single runtime. The core is written in C++ (raylet, GCS, object store, scheduling) with a Python frontend, communicating via gRPC and Arrow-compatible plasma object storage. Production entry points include `ray.init()` for cluster bootstrap, `@ray.remote` for task/actor offload, Ray Serve for model deployment (FastAPI/DAG/gang-scheduling patterns), Ray Data for streaming ETL, Ray Train/Tune for distributed training and hyperparameter search, and RLlib for reinforcement learning. The architecture is a hybrid: a distributed systems platform (scheduler, object store, autoscaler) layered with high-level AI libraries, all built via Bazel with extensive CI (Docker images, build matrices, release tests). It is not a data research repo, though Ray Data and RLlib contain research-grade data pipeline and RL primitives; the dominant character is a production-grade application framework.
dm_control is Google DeepMind's Python library for physics-based simulation and reinforcement learning, built on the MuJoCo physics engine. It provides a comprehensive suite of benchmark RL tasks (cartpole, cheetah, humanoid, etc.), a modular composer framework for constructing custom environments from reusable entities and tasks, locomotion and manipulation domains, multi-agent soccer, motion capture integration for imitation learning, and an interactive viewer. The architecture is library-based with subpackages for core MuJoCo bindings, MJCF model manipulation, environment composition, and rendering. Key entry points include suite.load() for standard tasks, composer.Environment for custom environments, and viewer.launch() for interactive visualization. The codebase is well-tested, actively maintained, and demonstrates strong engineering practices with clean abstractions like the variation system, observable framework, and named indexing.
OpenRLHF is a production-oriented distributed RLHF/RL training framework for large language models, not a research notebook collection. The codebase implements a full training stack: SFT, reward modeling, DPO/IPO/cDPO, and RL algorithms including PPO, REINFORCE++, GRPO, RLOO, Dr.GRPO, and FlashREINFORCE. Its architecture combines Ray for distributed actor orchestration and placement groups, vLLM for high-throughput rollout generation, DeepSpeed ZeRO for model training, and HuggingFace Transformers for model handling. The primary entry points are CLIs under openrlhf/cli, with train_ppo_ray.py as the main RL launcher; these accept a large set of hierarchically parsed dotted flags that control training, rollout, reward, vLLM, and DeepSpeed behavior. The repository is genuinely hybrid in spirit because it trains research-scale models and reproduces academic methods, but its dominant character is application/framework: it ships packaging, Dockerfiles, CLI entry points, unit tests, and versioned releases. The design emphasizes a token-in-token-out agent abstraction that unifies single-turn and multi-turn RLHF, plus sync and async training modes, a hybrid engine that colocates training and generation on shared GPUs, and VLM support. Distributed checkpointing, LoRA/QLoRA, packing, dynamic batch sizing, importance sampling corrections, and KL-estimator variants are all present in the code, not merely in the README. Craftsmanship is strong. The code separates concerns across cli/, trainer/, models/, datasets/, and utils/, and includes several genuinely reusable patterns such as a self-contained math answer grader, a flat-to-hierarchical argparse config converter, a typed RL experience dataclass, and distributed loss-normalization utilities. The test suite exercises loss estimators, trainer control flow, experience batching, and agent behavior with mocked heavy dependencies, which is a pragmatic way to test GPU-bound code on CPU. The main weaknesses are operational rather than architectural: dependencies are extremely tightly pinned (deepspeed==0.19.6, transformers==5.15.0, ray==2.55.0, vllm==0.29.0), and the code monkey-patches upstream internals and relies on version-specific behaviors. This makes upgrades brittle, and the CI pipeline builds wheels without running the test suite. Overall, this is a mature, actively maintained framework with high reusability for engineers building LLM alignment pipelines, though it demands GPU infrastructure and careful dependency alignment.
This repository is the official companion code for the book "Build a Reasoning Model (From Scratch)" by Sebastian Raschka (Manning, 2025). It is primarily a data_research codebase focused on building, evaluating, and training reasoning LLMs from scratch using the Qwen3 0.6B model as a base. The code covers inference-time scaling (temperature/top-p sampling, self-consistency), reinforcement learning from verifiable rewards (GRPO/RLVR), self-refinement, and knowledge distillation from larger teachers (DeepSeek-R1, Qwen3 235B). It includes a full custom Qwen3 implementation with KV cache, RoPE, and grouped-query attention, plus training scripts, notebooks, benchmark evaluations (MATH-500, MMLU), and Hugging Face export adapters. While educational in nature, it demonstrates production-quality engineering for research experiments. The repository is hybrid: it provides a reusable Python package for text generation and training, but its dominant character is data_research due to its notebook-driven, experimental, and benchmark-focused workflow.
Stable-worldmodel is a research framework for goal-conditioned world models in reinforcement learning, providing a full pipeline from data collection and storage to model training and planning-based evaluation. The codebase is primarily a research platform with strong software engineering practices—pluggable data formats (Lance, HDF5, video, folder), a command-line interface (swm), and a Hydra-based configuration system for reproducible experiments. It includes implementations of multiple world model architectures (LeWM, PLDM, PreJEPA/DINO-WM, TD-MPC2, GCRL), planning solvers (CEM, MPPI, iCEM, gradient-based, Lagrangian), and environment wrappers with domain randomization across 30+ tasks (DeepMind Control, Gymnasium, PushT, OGBench, Craftax, robotics). The repository demonstrates high craftsmanship in its data engineering: the data format registry extends cleanly, the Lance format supports cloud-native streaming from Hugging Face Buckets, and utilities like GoalDataset, HistoryBuffer, and ReplayBuffer solve recurring RL data management problems. The project is actively maintained with extensive test coverage, comprehensive documentation, and a scalable training pipeline using stable-pretraining and Lightning. While the primary goal is research (testing hypotheses about world model generalization under variation), the codebase includes application-level production quality—CI/CD, CLI, checkpoint management, and robust error handling—making it a rare hybrid that serves both researchers and engineers.
Checkpoint Engine is a production-grade Python library that serves as a lightweight middleware for updating model weights in large language model (LLM) inference engines such as vLLM and SGLang, a critical step in reinforcement learning fine-tuning. The system is designed to efficiently broadcast or peer-to-peer transfer sharded checkpoint data (safetensors format) across a cluster of GPUs, drastically reducing weight update time—for example, updating a 1-trillion-parameter model across thousands of GPUs takes approximately 20 seconds. The architecture is a distributed parameter server colocated with inference workers, using ZeroMQ for inter-process communication (IPC) and optionally Mooncake Transfer Engine for RDMA-based P2P transfers. It supports multiple accelerator backends (CUDA, NPU, XPU) through a unified device abstraction layer. The codebase is well-factored, with clear separation of concerns: IPC handling, memory pinning, distributed communication, and worker integration are each encapsulated in their own modules. The project includes an extensive test suite covering hardware-gated and CPU-only scenarios, CI/CD pipelines, and pre-commit linting. While tightly integrated with vLLM via a worker extension plugin, the core components (IPC handler, device manager, checkpoint loading) are designed for reuse and could be extracted for other inference frameworks.
Vowpal Wabbit is a production-ready machine learning system specializing in online learning algorithms, particularly contextual bandits. It features a modular reduction-based architecture where learning algorithms can be composed and stacked. The system provides a command-line interface, Python package, and bindings for C#/.NET, Java, and WebAssembly. Its core innovation lies in the reduction framework that allows flexible composition of learning algorithms, efficient feature hashing for scalable representation, and extensive support for various learning paradigms including supervised learning, multilabel classification, cost-sensitive learning, and reinforcement learning.
SGLang is a high-performance inference serving system for large language models (LLMs) and multimodal models (vision, video, audio, diffusion). It provides an OpenAI-compatible API, a Python frontend DSL for programmable inference, and supports a wide range of hardware backends (NVIDIA CUDA, AMD ROCm, Intel XPU, Ascend NPU, Apple MLX). The architecture is a distributed serving runtime (SRT) with a scheduler, model executor, custom CUDA/Triton kernels, and advanced features like speculative decoding (EAGLE, MTP, NGRAM), disaggregated prefill/decode, hierarchical KV cache (HiCache), and quantization (FP8, NVFP4, AWQ, GPTQ). It also includes a Rust-based model gateway for routing and a multimodal generation pipeline for diffusion models. The system is designed for production-scale deployment, powering trillions of tokens daily across hundreds of thousands of GPUs.
KataGo is a production-grade Go-playing AI engine that combines deep neural networks with Monte Carlo Tree Search (MCTS). The system supports self-play reinforcement learning, GTP protocol for GUI integration, analysis mode, and distributed training. Its architecture features a C++ core with multiple GPU backends (CUDA, ROCm, OpenCL, Metal, ONNX, TensorRT) for neural network inference, a comprehensive MCTS search implementation with graph enhancements, and Python training scripts. The codebase demonstrates strong engineering with well-abstracted backends, thread-safe utilities, and a modular configuration system. External dependencies include NVIDIA CUTLASS and cuDNN frontend for high-performance GPU kernels, indicating a focus on cutting-edge hardware utilization.
Qlib is an open-source, AI-oriented quantitative investment platform developed by Microsoft. It covers the full pipeline for quantitative research and trading: data ingestion and feature engineering (Alpha158, Alpha360, high-frequency), model training (traditional ML like LightGBM/XGBoost, deep learning in PyTorch including LSTM, GRU, Transformer, GATs, and reinforcement learning), backtesting with realistic costs and slippage, portfolio optimization, and online serving for production-like deployment. The platform is highly modular and config-driven, with a core library (qlib) and a separate contrib layer for community extensions. It includes a CLI (qrun) for running full workflows from YAML configs, extensive experiment tracking via MLflow, and a rich set of evaluation metrics. While primarily a library/application, it also functions as a research benchmark suite with many example models and data collectors for global markets. The codebase is active and well-maintained, though some components (e.g., TFT model, older RL code) show signs of technical drift due to dependency version locks and deprecated patterns.
This repository is a comprehensive educational curriculum for Machine Learning, structured as a collection of Jupyter notebooks, data files, and instructional documentation. While it contains small application examples (a Vue.js quiz app and a Flask web app demo), the dominant character is data research and education, focusing on teaching ML concepts through practical data science workflows. The content covers Regression, Classification, Clustering, NLP, Time Series, and Reinforcement Learning, utilizing Scikit-learn as the primary modeling library. The codebase is organized into modular lesson directories, each containing notebooks, assignments, and solutions. It demonstrates a strong commitment to accessibility with extensive multi-language translations and a devcontainer setup for consistent environments. However, reproducibility is impacted by a reliance on external datasets (Kaggle, Dropbox) which are prone to link rot, and scattered dependency management across different lesson folders rather than a unified root configuration. Technically, the repository shows signs of medium drift, particularly in the quiz application which utilizes Vue 2 patterns instead of the modern Vue 3 Composition API. The core ML content remains relevant, but the educational infrastructure (translation services, external links) requires ongoing maintenance to ensure accuracy and availability. Despite this, it stands as a high-quality, reusable resource for learning ML pipelines and data engineering patterns.
This repository is a comprehensive collection of annotated PyTorch implementations of deep learning papers, spanning transformers, GANs, reinforcement learning, diffusion models, and more. Each implementation includes side-by-side notes explaining the code and underlying concepts, making it an educational resource for understanding and reproducing paper results. The codebase is organized by paper/model type, with standalone experiment scripts, reusable modules, and configuration-driven training loops. Key strengths include clean abstractions for transformer components, a flexible configs system, and extensive coverage of both foundational and modern architectures. However, the project has not seen commits since early 2024, some dependencies are outdated, and certain experiments require large GPU resources or external data not bundled in the repo. While individual components are well-crafted and reusable, the overall collection is best used as a reference and learning tool rather than a production framework.
This repository is a structured, comprehensive AI engineering curriculum spanning 20 phases and over 500 lessons, from math foundations and ML fundamentals to deep learning, computer vision, NLP, speech, transformers, generative AI, reinforcement learning, LLMs from scratch, agent engineering, multi-agent systems, production infrastructure, and ethics/safety. It is an educational resource, not a production application or data research project. The code consists of self-contained pedagogical implementations (primarily Python, with some TypeScript, Rust, and Julia) that demonstrate core algorithms from scratch—such as attention mechanisms, BPE tokenization, diffusion models, RLHF, and agent loops—using synthetic data and deterministic stubs. Capstone projects include small web servers (Hono/TypeScript) and test harnesses. The curriculum is actively maintained with CI tooling, certification prep materials, and translation support. While the code is simplified for learning, it exhibits clean patterns, consistent lesson structure, and testing practices that make individual components (e.g., tool registries, state machines, evaluation harnesses) reusable in other contexts.
Luet is a container-native package manager for building, installing, and managing software packages. It uses Docker/buildah as build backends, treats packages as container images, and employs SAT solving (with optional Q-learning) for dependency resolution. The application features a CLI-driven architecture with commands for package operations, repository management, and system maintenance. Key architectural elements include a modular design with pluggable solvers, compiler backends, and repository clients, plus a context-based dependency injection system for logging, configuration, and garbage collection. The system stores package metadata in BoltDB or in-memory databases and supports templating via Sprig for dynamic package definitions.
This repository implements a complete reinforcement learning pipeline for whole-body motion tracking on humanoid robots, specifically Unitree G1 and XBot L7. The system uses a teacher-student architecture with Proximal Policy Optimization (PPO) to train policies that track reference motions from retargeted motion capture datasets. It features three training phases: a base training phase, an adaptation phase with domain randomization, and a finetuning phase. The environment is built on MJLab (a MuJoCo wrapper) and includes extensive domain randomization (e.g., motor parameters, body mass, center of mass, joint offsets, external forces, and a window load capacity curriculum for payload-aware training). The codebase is highly modular, with registered MDP components (observations, rewards, terminations, randomizations) and a flexible configuration system using Hydra. It also includes advanced features such as symmetry augmentation, differentiable forward kinematics, and a custom motion dataset pipeline with memory-mapped storage. While primarily a reinforcement learning training framework, it has strong data engineering aspects for processing motion datasets and generating training data for the student encoder. The repository is well-structured for research and sim-to-real transfer, but it is tightly coupled to specific robot assets and dependencies (torch, torchrl, mjlab) with pinned versions that may require adaptation for other environments.
This repository is an extensive, open-source educational tutorial on building and training large language models (LLMs) from scratch, maintained by the Datawhale community. It covers the full pipeline: NLP fundamentals, Transformer architecture (with a custom Tiny-K implementation), pretraining, supervised fine-tuning (SFT), parameter-efficient fine-tuning (LoRA), reinforcement learning (GRPO, OPD, Search-R1, ReTool), and practical applications like RAG and Agent systems. The code is organized into chapters with clear progression, and each chapter includes runnable scripts, configuration files, and detailed documentation in Chinese. While primarily educational, the repository contains several well-crafted, reusable components such as a modular Transformer implementation, efficient dataset loaders, a RAG pipeline, an Agent framework, and RL training loops with distributed support. The project is actively maintained with recent additions (Extra-Chapter) and has a large community following. However, reproducibility is moderate due to fragmented dependency files, hardcoded paths, and reliance on external services (e.g., DeepSeek Search, SwanLab). The code is not production-grade but serves as an excellent learning resource and a source of reusable patterns for LLM research and development.
GEM is a Gym-like framework for training and evaluating agentic LLMs via reinforcement learning. It provides a diverse set of environments (games, math, code, QA, etc.) that can be used to train LLMs to interact with tools, solve problems, and make sequential decisions. The framework features a modular design with composable wrappers for observation modification, tool integration, and episode tracking, along with vectorized environments for parallel execution. It supports integration with multiple RL training frameworks (Oat, Verl, Tinker, OpenRLHF, RL2) through example scripts, making it a versatile tool for LLM agent research.
This repository is a massive, multi-disciplinary research codebase from Google Research, primarily focused on machine learning, optimization, and computer science experiments. While it contains a few application-like components (e.g., a Next.js frontend for model exploration, a Rust-based 3D splatting renderer), the dominant character is data_research, hosting thousands of independent experimental projects. The codebase serves as a central hub for publishing reproducible research in domains including Reinforcement Learning, Computer Vision (NeRF, 3D), NLP (Transformers, LLMs), and Privacy. It relies heavily on shell scripts for environment setup and experiment orchestration, with a mix of Python, C++, Rust, and Java implementations. The infrastructure is robust for research (e.g., `scann` for search, `lila` for experiment management) but exhibits significant fragmentation and technical debt due to the rapid pace of research and mixed framework usage.
PhySO (Physical Symbolic Optimization) is a Python library for symbolic regression using deep reinforcement learning, with a strong emphasis on physics-informed constraints such as dimensional analysis. It discovers closed-form analytical expressions from data, supporting both single-dataset symbolic regression (SR) and multi-dataset class symbolic regression (Class SR) where the same functional form is shared across realizations with per-dataset free constants. The core architecture consists of a token-based program representation (reverse Polish notation), a custom LSTM-based policy network, a modular prior system (hard/soft length, physical units, nesting limits), and an efficient batch execution engine with optional parallel free-constant optimization. The library provides high-level entry points (`physo.SR()` and `physo.ClassSR()`), configuration presets, and a toolkit for expression manipulation and random generation. It includes extensive benchmarking infrastructure (Feynman and Class benchmarks), demos for real-world problems (e.g., Milky Way streams, US tariffs), and unit tests. External dependencies include PyTorch, NumPy, SymPy, pandas, scikit-learn, and matplotlib. While primarily a software library for symbolic regression, it is deeply research-oriented with benchmark result CSVs and analysis scripts, giving it a hybrid application/research character.