Awesome GitHub RepositoriesPreference Optimization

This project is a fine-tuning framework and training pipeline designed to optimize and adapt large language and vision models. It provides a specialized toolkit for parameter-efficient tuning and supervised learning, serving as both a trainer for multimodal models and a deployment tool for serving fine-tuned models via high-performance inference engines. The framework focuses on reducing memory and compute requirements by updating a small subset of model parameters. It supports a wide range of adaptation strategies, including vision-language model training to align text, image, video, and aud

This project provides a comprehensive framework for building, training, and managing autonomous agents. It enables the construction of systems that utilize language models to plan, manage memory, and execute multi-step tasks through iterative reasoning loops and tool-based actions. The framework distinguishes itself by offering specialized capabilities for interacting with graphical user interfaces and legacy software, allowing agents to perceive visual elements and perform actions like a human user. It supports complex, cross-application workflows through graph-based orchestration and provid

This project is a comprehensive framework for the entire lifecycle of transformer-based language models, supporting everything from foundational pretraining to specialized deployment. It provides a modular toolkit for defining neural network architectures, managing data preparation pipelines, and executing training routines across various scales. The framework is designed to handle the full model development process, including supervised fine-tuning, behavioral alignment, and the integration of agentic capabilities. What distinguishes this framework is its focus on efficient training and adva

This project is a comprehensive framework and toolkit for developing, optimizing, and deploying transformer-based models across multimodal, document intelligence, and natural language processing tasks. It provides a unified neural architecture that processes text, vision, audio, and document layout data through a shared set of weights, enabling researchers and developers to build foundational models that align cross-modal representations. The platform distinguishes itself through advanced training and inference strategies designed for large-scale deep learning. It incorporates specialized mec

This project is a distributed training infrastructure designed for aligning large language models through reinforcement learning. It functions as an end-to-end engine for complex alignment tasks, including proximal policy optimization, direct preference optimization, and iterative self-play. By providing a unified framework for multi-turn interactions and tool-use scenarios, it enables the development of models capable of reasoning and external environment engagement. The framework distinguishes itself through a decoupled architecture that separates model training from sample generation. This

This project is a transformer post-training toolkit and reinforcement learning library designed to align language model behavior with human preferences. It provides a framework for managing the transition from supervised fine-tuning to reinforcement learning and preference optimization. The library distinguishes itself through a specialized focus on preference optimization and reward modeling, enabling the adjustment of model outputs based on preferred versus rejected examples. It also includes capabilities for training agents within controlled sandbox environments using task suites and verif

This library provides a comprehensive framework for fine-tuning, aligning, and distilling transformer-based language models. It serves as a toolkit for adapting models to specialized domains through supervised learning, while offering advanced methodologies to improve output quality and reasoning capabilities. The project distinguishes itself through specialized alignment and optimization techniques, including direct preference optimization and reinforcement learning, which allow models to be tuned against human preferences without complex reward modeling. It further supports training efficie

Axolotl is a configuration-driven framework designed for the fine-tuning, evaluation, and quantization of large language models. It functions as a comprehensive orchestrator for distributed training, enabling users to manage complex workflows across multi-node and multi-GPU environments. By utilizing structured configuration files, the platform streamlines the setup of training parameters, dataset paths, and hardware distribution strategies. The project distinguishes itself through its support for diverse training methodologies, including full-parameter tuning, parameter-efficient adaptation,

Airllm is a framework designed to execute and fine-tune large language models on consumer-grade hardware. By employing layer-wise model decomposition and memory-efficient loading techniques, the engine enables the operation of massive models that would otherwise exceed available system or video memory. The project distinguishes itself through a suite of optimization strategies that balance memory footprint with performance. It utilizes block-wise weight quantization and asynchronous layer prefetching to reduce resource consumption and hide data transfer latency. Additionally, the framework su

InternVL is a vision-language model framework that fuses a visual encoder with a large language model to translate image features into textual tokens for reasoning. It provides a system for multimodal inference and dialogue, enabling the processing of images and text to answer questions or generate descriptions. The project is distinguished by its high-resolution image processing, which uses dynamic tiling to maintain detail for images up to 4K resolution, and its chain-of-thought visual reasoning for solving complex mathematical and spatial problems. It also supports temporal frame sampling

OpenRLHF is a training framework and alignment library designed for reinforcement learning from human feedback across distributed GPU clusters. It provides tools for aligning large language models and multimodal vision-language models using algorithms such as PPO, GRPO, and DPO. The framework distinguishes itself through a distributed inference engine that overlaps sample rollout with training to increase throughput. It supports scaling to models exceeding 70 billion parameters via parameter sharding and handles long-context sequences through ring-attention sequence parallelism. The project

Intel XPU LLM Acceleration Library is a toolkit designed to accelerate large language model inference and finetuning on Intel CPUs, GPUs, and NPUs. It provides a distributed inference engine for scaling models across multiple accelerators, a multimodal model runtime for vision and speech tasks, and a low-bit model quantization tool for converting weights into INT4, FP8, and GGUF formats. The project features a parameter-efficient finetuning framework that enables model adaptation using QLoRA and DPO on Intel hardware. It distinguishes itself by providing specialized optimizations for Intel XP

This project is an educational program focused on the alignment of small language models. It provides a technical curriculum and a series of courses designed to teach how to align models with human preferences and behaviors. The material covers the implementation of preference optimization algorithms and the adaptation of vision-language models to process both text and image data simultaneously. It also includes instructional guides on synthetic data generation to improve model performance in specialized domains. The curriculum encompasses supervised fine-tuning workflows, the use of chat te

SkyReels-V2 is a video generation system that creates, extends, and refines video clips from text descriptions, images, or both. It operates as a diffusion-based video generation model that can produce videos of any duration by denoising frames sequentially, with each new frame conditioned on the ones that came before it. The system supports generating videos from scratch using text prompts, starting from a single image and producing subsequent frames, or constraining both the first and last frames to match user-provided images. What distinguishes SkyReels-V2 is its combination of infinite-le

Liger-Kernel is a collection of pre-built fused Triton kernels and patching utilities designed to accelerate large language model training. It provides drop-in kernel replacements for common LLM operations such as RMSNorm, cross-entropy loss, and attention, enabling increased throughput and reduced memory usage while preserving bitwise-exact gradients. The project serves as a toolkit for composing custom model architectures from individual optimized kernels and for patching pre-existing models with minimal code changes. The project distinguishes itself through its ability to perform runtime m

Torchtune is a PyTorch-native library for fine-tuning, aligning, and quantizing large language models. It provides a config-driven system for instantiating components, orchestrating distributed training, and managing parameter-efficient fine-tuning with quantization support, all through YAML-based configurations and command-line overrides. The library distinguishes itself through its comprehensive post-training workflow orchestration, combining supervised fine-tuning, preference optimization (DPO, PPO, GRPO), knowledge distillation, and quantization-aware training in a single configurable pip

xtuner ist eine umfassende Trainings-Engine für Large Language Models und bietet ein Toolkit für Pre-Training, Supervised Fine-Tuning und die Optimierung von vision-sprachlichen multimodalen Modellen. Sie dient als verteilter Trainingsbeschleuniger und spezialisiertes Framework zur Skalierung von Mixture-of-Experts-Modellen sowie zur Ausrichtung von Modellverhalten durch Reinforcement Learning from Human Feedback. Das Projekt zeichnet sich durch fortgeschrittene Speicher- und Rechenoptimierungen aus, wie Sequence-Parallelism für ultra-lange Kontextfenster und Interleaved-Pipeline-Parallelism zur Reduzierung von GPU-Idle-Zeiten. Es bietet eine dedizierte Suite für Preference-Optimization und implementiert Techniken wie Group Relative Policy Optimization und Direct Preference Optimization, um Modell-Policies und Belohnungssysteme zu verfeinern. Breite Funktionsbereiche decken verteiltes Modelltraining über mehrere Knoten hinweg, multimodale Datensatzvorbereitung und die Verwaltung von Adapter-basiertem Fine-Tuning ab. Die Engine enthält zudem Tools für Modellevaluation, Weight-Merging und den Export trainierter Parameter in Inferenz-Engines. Das Training wird über standardisierte Konfigurationsdateien und verteilte Launcher verwaltet, um konsistente Ergebnisse über Rechencluster hinweg sicherzustellen.

LLM-RL-Visualized ist eine visuelle Referenzbibliothek und eine Sammlung von Wissenskarten, die Large Language Model- und Reinforcement Learning-Algorithmen erklären. Es bietet ein strukturiertes System aus konzeptionellen Diagrammen und Taxonomien, das die Schnittmenge von Sprachmodell-Alignment und Reinforcement Learning abdeckt. Das Projekt zeichnet sich durch detaillierte visuelle Mappings komplexer Workflows aus, wie etwa die Koordination von Reward-Modellen und Policy-Optimierung beim Reinforcement Learning from Human Feedback (RLHF). Es stellt verschiedene Preference-Optimization-Architekturen gegenüber, wie RLHF und Direct Preference Optimization, und zeichnet die theoretische Abstammung von Reinforcement-Learning-Algorithmen von Markov-Entscheidungsprozessen bis hin zu Actor-Critic-Frameworks nach. Die Bibliothek deckt ein breites Spektrum an Funktionen ab, darunter LLM-Inferenzoptimierung, parameter-effiziente Fine-Tuning-Techniken und die sequenziellen Phasen der Modellentwicklungspipeline. Sie bietet zudem strukturelle Diagramme für Modellkonfigurationen, Visualisierungen von Token-Decoding-Strategien sowie operative Abläufe für Retrieval-Augmented Generation und Tool-Integration. Zusätzliche Inhalte umfassen Illustrationen grundlegender neuronaler Netzwerkoperationen und logischer Schlussmechanismen wie Monte Carlo Tree Search und Knowledge Distillation.

LLaDA is a masked diffusion language model and conditional text generator. It generates text by iteratively refining masked tokens through a diffusion process rather than predicting the next token in a sequence. The project functions as a vision-language diffusion model, converting visual inputs into text responses. It also serves as a preference optimization framework that uses log-likelihood estimation and evidence lower bounds to tune model responses. The system supports multi-round conversational AI and text sequence evaluation. It integrates vision-language embedding for cross-modal con

Tinker Cookbook is an open-source framework for fine-tuning large language models, supporting supervised learning, reinforcement learning, and parameter-efficient techniques like LoRA adapters. It provides a complete pipeline for aligning models with human preferences through multi-stage RLHF workflows, from supervised fine-tuning through preference optimization to reinforcement learning. The framework distinguishes itself through recipe-based training orchestration, where fine-tuning workflows are defined as composable recipe files that chain data loading, model configuration, and training l