Coding & Development

simple_GRPO is an open-source implementation of the GRPO algorithm, specifically designed for reproducing r1-like LLM thinking. It utilizes a core loss calculation formula referenced from Hugging Face's trl, but with a significantly simplified codebase. The tool aims to save GPU memory, enabling feasible and efficient training, and helps users quickly understand and experiment with Reinforcement Learning processes like GRPO. It supports features such as improved multi-answer generation, regrouping, penalty on KL, and parameter tuning, all within approximately 200 lines of code across two files. The reference model is decoupled, allowing it to run on separate GPUs, which prevents multiple copies from being created by torch’s multiprocessing and enables training of large models on less powerful hardware.

tiny-diffusion offers a character-level language diffusion model for text generation, implemented in just 365 lines of Python code. This compact model, with 10.7 million parameters, is trained on Tiny Shakespeare, making it suitable for local experimentation and learning. The repository also features a tiny GPT implementation in 313 lines, with significant code overlap between the two models. It supports parallel decoding for diffusion and autoregressive generation for GPT. Users can train both models from scratch, visualize the generation process, and compare the diffusion and GPT models side-by-side. The diffusion model introduces key modifications like a mask token, bidirectional attention, confidence-based parallel decoding, and a training objective focused on unmasking.

tomesd is an open-source Python and PyTorch-based tool designed to accelerate Stable Diffusion models by implementing Token Merging (ToMe). This technique reduces computational load by merging redundant tokens within the transformer blocks, leading to faster image generation and lower memory consumption. tomesd works out-of-the-box with various Stable Diffusion models, including v1, v2, Latent Diffusion, and Diffusers, and does not require additional training. While it's a lossy process, it minimizes quality degradation while providing substantial speed and memory benefits. It can be applied to existing Stable Diffusion environments and is compatible with other efficient transformer implementations like xformers.

LENS Corporation specializes in developing custom AI-powered solutions with a strong focus on biometrics and computer vision. They offer state-of-the-art SDKs for face, fingerprint, and iris recognition, enabling real-time automated biometric applications even on edge devices without an internet connection. Beyond biometrics, LENS provides advanced image analysis services, allowing businesses to outsource complex visual data processing to intelligent, adaptively learning machines. The company also excels in cross-media translation, delivering solutions like text-to-speech, speech-to-text, and image-to-text to enhance business convenience. LENS emphasizes ethical AI development, ensuring their solutions are explainable, transparent, and compliant with data privacy regulations like GDPR and HIPAA, while granting full intellectual property rights to their clients.

ThinkGPT is a Python library designed to augment Large Language Models (LLMs) by implementing Chain of Thoughts techniques. It enables LLMs to think, reason, and act as generative agents, addressing common limitations such as restricted context windows. Key features include memory management for LLMs to recall past experiences, self-refinement capabilities to improve model-generated content, and knowledge compression techniques to fit extensive information within an LLM's context. The library also offers inference based on available data, natural language conditions for decision-making, and efficient context length management, all through an easy-to-use Pythonic API.

Trankit is a light-weight, transformer-based Python toolkit designed for multilingual Natural Language Processing (NLP). It offers a trainable pipeline for fundamental NLP tasks across more than 100 languages, and includes 90 downloadable pretrained pipelines for 56 languages. Trankit outperforms other state-of-the-art multilingual toolkits like Stanza in various tasks, including sentence segmentation and dependency parsing, while maintaining efficiency in memory usage and speed. Key features include an Auto Mode for automatic language detection, a command-line interface for ease of use, and support for tasks such as tokenization, part-of-speech tagging, morphological feature tagging, dependency parsing, and named entity recognition. It also allows users to build and share customized pipelines.

WideLabs specializes in delivering sovereign AI infrastructure tailored for businesses. The platform provides robust GPU cloud services, enabling companies to run demanding AI workloads efficiently. Beyond infrastructure, WideLabs also develops and integrates proprietary AI models, offering advanced capabilities for various business needs. Their end-to-end solutions ensure comprehensive support from deployment to ongoing management, addressing complex challenges in generative AI, computer vision, and predictive algorithms. WideLabs aims to create a significant impact on individuals, institutions, and companies by leveraging cutting-edge AI technologies.

Unsloth is an open-source platform designed for training and running a wide array of open models, including Gemma 4, Qwen3.5, DeepSeek, and gpt-oss, directly on local machines. It offers a user-friendly web UI, Unsloth Studio, for easy interaction, alongside a code-based version, Unsloth Core. The tool boasts significant performance improvements, enabling up to 2x faster training with up to 70% less VRAM, without compromising accuracy. It supports various model types including text, audio, embedding, and vision models, and provides features like model inference, export, tool calling, and code execution. Unsloth also includes advanced training capabilities such as reinforcement learning, custom Triton kernels, and data recipes for dataset creation from diverse file types.

The uvadlc_notebooks repository offers a comprehensive collection of Jupyter notebook tutorials specifically designed for the Deep Learning Course at the University of Amsterdam (MSc AI). These notebooks aim to bridge the gap between theoretical concepts and practical implementation, covering diverse topics such as optimization techniques, transformers, graph neural networks, and more. The materials are available for both Fall 2023 and Fall 2024 course editions, with support for PyTorch and PyTorch Lightning, as well as JAX+Flax. Users can run the notebooks locally on CPU, utilize Google Colab for GPU access, or leverage the Snellius cluster for larger-scale training. The tutorials are integrated into PyTorch Lightning's official documentation, making them a valuable resource for students and practitioners alike.

OpenManus-RL is an open-source initiative, collaboratively led by Ulab-UIUC and MetaGPT, dedicated to advancing reinforcement learning (RL) tuning for large language model (LLM) agents. Inspired by successful RL tuning in models like Deepseek-R1, this project explores novel algorithmic structures, diverse reasoning paradigms, and sophisticated reward strategies. It supports rigorous testing on agent benchmarks such as GAIA, AgentBench, WebShop, and OSWorld, with all progress and tuned models openly shared. The platform integrates advanced RL algorithms like PPO and DPO through the Verl submodule, offering efficient and flexible training capabilities. It also provides a simplified library for Supervised Fine-Tuning (SFT) and GRPO tuning, making it a comprehensive solution for researchers and developers looking to push the boundaries of agent reasoning and tool integration.

ai-agents-masterclass is a comprehensive GitHub repository designed to accompany an AI Agents Masterclass video series. It offers all the code and resources used in the YouTube series, enabling developers to follow along and build their own AI agents. The masterclass focuses on empowering developers to leverage AI agents for transforming businesses and creating sophisticated software. The repository includes examples for building agents with LangChain, LangGraph, n8n, and other technologies, covering topics from basic agent creation to RAG agents, task management, and deployment. It serves as a practical guide for anyone looking to dive deep into AI agent development.

Jynnt is an AI paradise designed to simplify and elevate your AI experience by offering a light, organized, and efficient workspace. It provides access to over 100 stellar AI models, allowing for limitless choices and multi-model interactions. Users can organize their chats with folders, spaces, and tags, eliminating clutter and enhancing productivity. The platform boasts a clean, minimal, and lightweight interface, ensuring a smooth and efficient workflow. Jynnt offers unlimited chats and messaging, along with 24/7 support and free updates, all under a Pay-As-You-Go pricing model.

AdderNet is an innovative AI framework designed to significantly reduce computation costs in deep neural networks, particularly convolutional neural networks (CNNs), by replacing traditional multiplications with more efficient additions. This is achieved by using the L1-norm distance between filters and input features as the output response. The framework demonstrates impressive performance, achieving 74.9% Top-1 accuracy and 91.7% Top-5 accuracy using ResNet-50 on the ImageNet dataset, all without any multiplication operations in the convolution layer. It also shows strong classification results on CIFAR-10 and CIFAR-100 datasets, as well as competitive super-resolution and adversarial robustness benchmarks. The project provides code for training and evaluation on these datasets, making it a valuable resource for researchers and developers focused on efficient deep learning.

adk-go is an open-source, code-first Go toolkit designed for building, evaluating, and deploying sophisticated AI agents with flexibility and control. It provides a modular framework that applies software development principles to AI agent creation, simplifying the orchestration of agent workflows from simple tasks to complex systems. While optimized for Gemini, ADK is model-agnostic and deployment-agnostic, ensuring compatibility with various frameworks. This Go version is particularly suited for developers creating cloud-native agent applications, capitalizing on Go's inherent strengths in concurrency and performance. Key features include idiomatic Go design, a rich tool ecosystem for diverse agent capabilities, and strong support for containerization and deployment in environments like Google Cloud Run.

AdvBox is a comprehensive open-source toolbox designed to generate adversarial examples that can fool neural networks across various popular machine learning frameworks, including PaddlePaddle, PyTorch, Caffe2, MxNet, Keras, and TensorFlow. Beyond generation, it also provides capabilities to benchmark the robustness of machine learning models. A key feature is its command-line interface, which allows users to generate adversarial examples with "Zero-Coding," making it accessible for those who prefer not to delve deep into programming. AdvBox is part of the broader Advbox Family, a suite of AI model security tools from Baidu Open Source, which also includes tools for detection and protection against adversarial examples, as well as attack and defense cases for different AI applications.

Adversarial-Attacks-PyTorch, also known as Torchattacks, is a comprehensive PyTorch library designed for generating adversarial examples to test the robustness of deep learning models. It provides a PyTorch-like interface, making it intuitive for developers and researchers already familiar with PyTorch to implement various adversarial attacks. The library supports a wide array of attacks, including FGSM, PGD, CW, AutoAttack, and many more, covering different distance measures like Linf, L2, L1, and L0. It also includes features for targeted attacks, saving/loading adversarial images, and managing model training modes during attacks. The tool emphasizes ease of use and performance, often outperforming other libraries in speed for certain attacks.

agentic_coding_flywheel_setup is an open-source tool designed to quickly set up a multi-agent AI development environment on a fresh Ubuntu VPS. In just 30 minutes, it configures essential components including coding agents, session management, safety tools, and coordination infrastructure. This tool is ideal for developers looking to rapidly deploy a fully configured agentic coding VPS, transforming a standard Ubuntu server into a powerful AI-powered development hub. It streamlines the setup process, allowing users to focus on AI development rather than environment configuration.

AnglE is an open-source library designed for training and inferring state-of-the-art BERT/LLM-based sentence embeddings. It utilizes an angle-optimized approach, offering various loss functions like AnglE loss, Contrastive loss, CoSENT loss, and Espresso loss. The library supports both BERT-based and LLM-based models, including bi-directional LLMs, and facilitates single-GPU and multi-GPU training. AnglE has achieved SOTA performance on benchmarks like STS and MTEB, with models trained using AnglE reaching top positions. It provides a flexible framework for researchers and developers to build and deploy high-quality sentence embedding models.

This open-source project offers a comprehensive collection of PyTorch implementations for various deep learning papers and algorithms. Each implementation is meticulously documented with side-by-side explanations, designed to enhance understanding of complex concepts. The repository is actively maintained, with new implementations added almost weekly, covering a wide range of topics including transformers (original, XL, Switch, Feedback, ViT), optimizers (Adam, AdaBelief, Sophia), GANs (CycleGAN, StyleGAN2), and reinforcement learning (PPO, DQN). It also features implementations for Capsule Networks, Distillation, and various normalization layers. This resource is ideal for students, researchers, and developers looking to delve deeper into the practical application and theoretical underpinnings of deep learning.

AlpacaEval is an automatic evaluator designed for instruction-following language models, providing a fast, cheap, and highly correlated alternative to human evaluation. It boasts a Spearman correlation of 0.98 with ChatBot Arena, costing less than $10 of OpenAI credits and running in under 3 minutes. The tool offers precomputed leaderboards for common models, an automatic evaluator validated against 20K human annotations, and a toolkit for building advanced automatic evaluators with features like caching, batching, and multi-annotators. It also includes 20K human evaluation data and a simplified AlpacaFarm evaluation dataset. AlpacaEval is particularly useful for rapid model development and iterative testing, though it cautions against replacing human evaluation for high-stakes decision-making due to potential biases and limitations in instruction representativeness.

Auto-Deep-Research is an open-source, fully-automated personal AI assistant designed as a cost-effective alternative to OpenAI's Deep Research. Built on the AutoAgent framework, it boasts high performance on the GAIA Benchmark and offers universal LLM support, seamlessly integrating with a wide range of models including OpenAI, Anthropic, Deepseek, vLLM, Grok, and Huggingface. The tool supports both function-calling and non-function-calling interaction LLMs and handles file uploads for enhanced data interaction. Users can get started instantly with a simple command, requiring zero configuration for an out-of-the-box experience. It aims to provide a personal assistant at a much lower cost, leveraging pay-as-you-go LLM API keys.

AutoCoder is an advanced AI model specifically designed for code generation tasks. It boasts impressive accuracy, surpassing GPT-4 Turbo (April 2024) and GPT-4o on the HumanEval base dataset. A key differentiator of AutoCoder is its innovative code interpreter, which automatically installs necessary packages and iteratively runs the generated code until it's deemed issue-free. This feature significantly expands the utility of the code interpreter compared to other models that may not access external libraries or run all generated code. AutoCoder is available in several model sizes, including AutoCoder (33B), AutoCoder-S (6.7B), and AutoCoder_QW_7B, with base models like deepseeker-coder and CodeQwen1.5-7b. It provides quick start guides for testing performance on benchmarks like HumanEval, MBPP, and DS-1000, and offers a web demo for interactive use.

albert_zh is an open-source implementation of A Lite Bert for Self-Supervised Learning of Language Representations, specifically optimized for Chinese language processing. Based on the BERT architecture, ALBERT introduces improvements like factorized embedding parameterization and cross-layer parameter sharing, significantly reducing the number of parameters while retaining or even improving accuracy. This leads to faster training and inference times, making it suitable for real-time applications and resource-constrained environments. The repository provides various pre-trained ALBERT models for Chinese, including tiny, small, base, large, and xlarge versions, with options for TensorFlow, PyTorch, and Keras. It includes scripts for pre-training on custom data and fine-tuning on downstream tasks like semantic similarity prediction, with examples provided for the LCQMC dataset.

ASearcher is an open-source framework designed for large-scale online reinforcement learning (RL) training of search agents, aiming to advance Search Intelligence to expert-level performance. It provides model weights, detailed training methodologies, and data synthesis pipelines, making it fully committed to open-source development. Key features include a prompt-based LLM agent for autonomous QA pair generation, a fully asynchronous agentic RL framework that decouples trajectory collection from model training, and the ability to enable long-horizon search with tool calls exceeding 100 rounds. ASearcher achieves cutting-edge performance on challenging QA benchmarks like GAIA, xBench-DeepSearch, and Frames, demonstrating substantial improvements through RL training. It also offers comprehensive guidance for building and training customized agents.