Coding & Development

HEBO is an open-source library developed by Huawei Noah's Ark Lab, focusing on Bayesian optimization, reinforcement learning, and generative model research. It offers official implementations for a wide range of algorithms, including Heteroscedastic Evolutionary Bayesian Optimisation (HEBO), a framework for Combinatorial and Mixed-variable Bayesian Optimization (MCBO), and End-to-End Meta-Bayesian Optimisation with Transformer Neural Processes (NAP). The library also covers high-dimensional Bayesian optimization with random decompositions (RDUCB) and applications in antibody design (AntBO) and logic synthesis (BOiLS). Additionally, HEBO supports research in reinforcement learning, such as enhancing agents with local guides and safe reinforcement learning, and generative models like EM-LLM for episodic memory in LLMs. It serves as a comprehensive resource for researchers and developers in these advanced AI fields.

Hedwig is an open-source repository offering PyTorch deep learning models specifically designed for document classification tasks. Developed by the Data Systems Group at the University of Waterloo, it includes implementations of several prominent models such as DocBERT, Reg-LSTM, XML-CNN, HAN, Char-CNN, and Kim CNN. Each model directory contains a detailed README.md for further information. The project is designed for Python 3.6 and PyTorch 0.4, with clear instructions for environment setup using Anaconda and installation of dependencies. It also provides options for downloading necessary datasets like Reuters, AAPD, and IMDB, along with word2vec embeddings, making it a comprehensive resource for document classification research and application.

GPTeacher is a comprehensive collection of modular datasets, meticulously generated by GPT-4, designed to facilitate various AI training and development tasks. The collection includes several distinct datasets: General-Instruct, Roleplay-Instruct, Code-Instruct, and Toolformer. The General-Instruct dataset, comprising approximately 20,000 examples, focuses on diverse tasks such as Chain of Thought Reasoning, Logic Puzzles, and Wordplay. The Roleplay-Instruct dataset, now in its V2 (Supplemental) version, is 2.5 times larger than the original and features simulated conversations for character role-playing. The Code-Instruct dataset offers around 5,350 code task instructions across various programming languages. Additionally, the Toolformer dataset is designed for training models to use predefined tools like search, Python, and Wikipedia. All datasets are formatted to be compliant with Alpaca's dataset structure, including instruction, input, and output fields, making them easy to integrate into existing fine-tuning processes.

GPTQ provides an efficient, open-source implementation of the GPTQ algorithm for accurate post-training quantization of generative pretrained transformers. This tool enables developers to compress large language models from the OPT and BLOOM families down to 2, 3, or 4 bits, significantly reducing their memory footprint and computational requirements while maintaining accuracy. Key features include support for weight grouping, evaluation of perplexity on various language generation tasks, and performance evaluation on ZeroShot tasks. The repository also offers a 3-bit quantized matrix full-precision vector product CUDA kernel and benchmarking code for individual matrix-vector products and language generation with quantized models. Recent updates include static groups options, adjusted preprocessing for C4 and PTB, optimized 3-bit kernels for faster generation, and a minimal LLaMa integration with new tricks like `--act-order` and `--true-sequential` for improved accuracy.

Hunyuan-A13B is an innovative and open-source large language model (LLM) developed by Tencent Hunyuan, featuring a fine-grained Mixture-of-Experts (MoE) architecture. With 80 billion total parameters and only 13 billion active parameters, it delivers high performance while maintaining optimal resource efficiency. Key features include hybrid reasoning support with both fast and slow thinking modes, ultra-long context understanding up to 256K tokens, and enhanced agent capabilities. The model is optimized for efficient inference using Grouped Query Attention (GQA) and supports multiple quantization formats like FP8 and INT4, making it suitable for resource-constrained environments. It is ideal for researchers and developers seeking powerful yet computationally efficient AI solutions.

Heretic is an open-source tool designed for the fully automatic removal of censorship, also known as "safety alignment," from transformer-based language models. It achieves this without requiring expensive post-training processes, utilizing an advanced implementation of directional ablation combined with a TPE-based parameter optimizer powered by Optuna. This approach allows Heretic to automatically find high-quality ablation parameters by co-minimizing refusal rates and KL divergence from the original model, ensuring the decensored model retains as much original intelligence as possible. The tool supports most dense and many multimodal models, including various MoE architectures. It also offers research features for interpretability studies, such as plotting residual vectors and printing residual geometry details.

hls4ml is an open-source Python package designed for machine learning inference on Field-Programmable Gate Arrays (FPGAs). It facilitates the creation of firmware implementations of machine learning algorithms using high-level synthesis (HLS) languages. The tool translates models from popular open-source machine learning frameworks, such as Keras, into HLS code, which can then be configured for specific use cases. While it originated from high-energy physics applications like L1 trigger systems at CERN, hls4ml has found diverse applications in areas such as quantum computing control systems, nuclear fusion feedback loops, low-power environmental monitoring on satellites, and biomedical signal processing. It supports various HLS backends including Xilinx Vivado HLS, Vitis HLS, Intel HLS, and Catapult HLS, with experimental support for Intel oneAPI.

HostedGPT is a free, open-source alternative to ChatGPT, built as a Ruby on Rails application, allowing it to be hosted anywhere or run locally. It supports multiple AI providers including Anthropic, Google, Llama, and Groq, enabling users to switch assistants mid-conversation. The platform offers a polished interface with strong mobile support and German localization. Users only pay for their API usage from providers like OpenAI, Anthropic, and Google, as the HostedGPT app itself is free. It also helps users avoid common usage caps and provides features for collecting, searching, and sharing conversations across different providers. Deployment options include Render, Fly.io, Heroku, or self-hosting, with detailed instructions for each.

Hands-On-Machine-Learning-with-CPP is a comprehensive code repository accompanying a Packt publication, designed to guide users through implementing various machine learning and deep learning algorithms using C++. It covers fundamental to advanced concepts, offering practical, easy-to-follow examples. Users will learn to preprocess diverse data types, employ key machine learning algorithms with C++ libraries, and optimize models using grid-search. The repository also includes methods for anomaly detection, improving collaborative filtering, and managing model structures. It provides a C++ program for image classification tasks with LeNet architecture, making it suitable for data analysts, data scientists, and machine learning developers looking to implement models in production.

MOFA-Video is an open-source project presented at ECCV 2024, designed for controllable image animation. It leverages generative motion field adaptions within a frozen image-to-video diffusion model to animate still images. The tool supports diverse control signals, including trajectories, keypoint sequences, and hybrid combinations, allowing for precise manipulation of motion. It features a sparse-to-dense motion generation approach and flow-based motion adaptation. MOFA-Video provides training scripts for trajectory-based and keypoint-based facial image animation, along with Gradio inference code and checkpoints for hybrid controls. This makes it a powerful resource for researchers and developers interested in advanced video generation techniques.

gpt-load is a robust, enterprise-grade AI API transparent proxy service built with Go, designed for developers and enterprises integrating multiple AI services. It features intelligent key management, including group-based management, automatic rotation, and failure recovery, ensuring high availability. The service supports weighted load balancing across multiple upstream endpoints and smart failure handling with automatic key blacklisting. It offers dynamic configuration with hot-reload capabilities, an enterprise-grade architecture supporting distributed leader-follower deployment, and a modern Vue 3-based web management interface. Comprehensive monitoring provides real-time statistics and detailed request logging, all optimized for high-concurrency production environments with zero-copy streaming and connection pool reuse.

GraphRAG Local Ollama is an open-source adaptation of Microsoft's GraphRAG, designed to leverage local models via Ollama for LLM and embedding extraction. This tool eliminates the dependency on costly OpenAPI models, offering a cost-effective solution for knowledge graph implementations. It supports a variety of local models such as Llama3, Mistral, Gemma2, and Phi3, and integrates with Ollama for both language models and embedding models like nomic-embed-text. The setup process is straightforward, involving conda environment creation, Ollama installation, repository cloning, and specific `pip install` commands. Users can easily configure models and run indexing and querying operations, with options to visualize generated graphs using tools like Gephi or a provided Python script.

ICEdit is an innovative open-source image editing tool that leverages a single LoRA (Low-Rank Adaptation) to achieve state-of-the-art instruction-based editing. It stands out by requiring only 0.5% of the training data and 1% of the parameters compared to prior SOTA methods, yet delivers fantastic image editing results. A key differentiator is its superior ID persistence, even surpassing models like GPT-4o. The tool is highly accessible, needing only 4GB VRAM to run, making it suitable for a wider range of hardware. ICEdit supports multi-turn and single-turn edits with high precision and offers various integration options, including official ComfyUI workflows and a Gradio demo for user-friendly interaction. It also provides training code for users to create their own editing LoRAs.

IDM-VTON is an open-source project that implements a novel approach to improving diffusion models for authentic virtual try-on in the wild. Based on research presented at ECCV 2024, this tool allows users to generate realistic virtual try-on images by integrating advanced diffusion techniques. It supports datasets like VITON-HD and DressCode, offering functionalities for both training and inference. The project provides detailed instructions for data preparation, model training, and running local Gradio demos, making it accessible for researchers and developers interested in virtual try-on technology.

iFlow CLI is a powerful AI assistant designed to run directly within your terminal, offering comprehensive command-line intelligence. It excels at analyzing code repositories, executing coding tasks, and interpreting user needs across various contexts. The tool significantly boosts productivity by automating a wide range of operations, from basic file management to intricate workflow automation. Key features include access to free AI models like Kimi K2 and Qwen3 Coder, flexible integration with existing development tools, and natural language interaction that eliminates the need for complex commands. It also supports advanced functionalities like SubAgents, custom commands, plan mode, and task tools, making it a versatile solution for developers seeking to streamline their workflow and enhance coding efficiency.

Open LLM Leaderboard is a comprehensive platform designed for tracking, ranking, and evaluating open-source large language models (LLMs) and chatbots. Hosted on Hugging Face Spaces, this tool provides a centralized hub for comparing the performance of different models across a range of standardized tests. Users can explore benchmarks such as IFEval, BBH, MATH, GPQA, MUSR, and MMLU-PRO, gaining insights into how various LLMs stack up against each other. The platform is particularly valuable for AI researchers and practitioners who need to assess model capabilities, identify top-performing models, and stay updated on the latest advancements in the open-source LLM landscape. While the live website content indicates a runtime error, the tool's core purpose is to offer transparent and data-driven evaluations.

Guidellm is an open-source platform designed for evaluating and enhancing Large Language Model (LLM) deployments, focusing on real-world inference needs. It simulates end-to-end interactions with OpenAI-compatible and vLLM-native servers, generating workload patterns that reflect production usage. The platform produces detailed reports to help teams understand system behavior, resource needs, and operational limits. Guidellm supports both real and synthetic multimodal datasets, including text, image, audio, and video inputs, and offers flexible execution profiles. It provides SLO-aware benchmarking, capturing complete latency and token-level statistics for metrics like TTFT, ITL, and end-to-end behavior, ensuring consistent assessment of model performance, tuning deployments, and capacity planning.

infiAgent, also known as MLA (Multi-Level Agent), is an open-source agent framework designed for handling long-running, complex tasks without issues like tool calling chaos or system crashes due to cumulative task resources and conversation history. It enables users to build powerful general-purpose and semi-specialized agents by simply editing configuration files. Key features include support for days-long complex tasks with full recovery from interruptions, compatibility with the Agent Skills open standard for dynamic skill loading, and a flexible architecture supporting both multi-level hierarchy and flat designs. The framework utilizes a file-directory-based memory system for persistent memory across sessions, eliminating the need for external databases. It also offers a Docker-based Web UI for multi-user registration and account management, and supports multi-provider model configurations for fine-grained cost control.

InstaFlow is an ultra-fast, one-step image generator that leverages Rectified Flow technique to achieve image quality comparable to Stable Diffusion while significantly reducing computational demands. It offers ultra-fast inference, generating images in approximately 0.1 seconds on an A100 GPU, saving about 90% of the inference time compared to original Stable Diffusion. InstaFlow generates high-quality images with intricate details and is compatible with pre-trained LoRAs and ControlNets. The training process is simple and efficient, involving supervised training and taking 199 A100 GPU days to train InstaFlow-0.9B. The tool provides code, pre-trained models, and a Hugging Face demo for easy access.

Archimyst is an industrial-grade coding CLI designed to optimize development workflows by providing a high-performance agentic runtime. It leverages specialized agent skills and precise architectural context to significantly reduce token usage, claiming up to a 90% saving. This tool is built for developers seeking to enhance efficiency and performance in their coding processes, particularly in managing complex system architectures. By offering a robust command-line interface, Archimyst integrates seamlessly into existing development environments, enabling more efficient code generation, simulation, and validation of production systems. Its focus on token economy makes it a valuable asset for cost-conscious development teams.

Improved-diffusion is an open-source codebase developed by OpenAI for working with Improved Denoising Diffusion Probabilistic Models. This repository provides the necessary tools and scripts for researchers and developers to train and sample from these powerful generative AI models. Users can prepare their own image datasets, including options for class-conditional training by naming files with labels. The codebase supports various hyperparameters for model architecture, diffusion processes, and training flags, allowing for flexible experimentation. It also facilitates distributed training across multiple GPUs and offers different sampling strategies, including DDIM. Pre-trained model checkpoints and their corresponding hyperparameters are provided for several common tasks, such as unconditional ImageNet-64 and CIFAR-10 generation, class-conditional ImageNet-64, and LSUN bedroom models.

IntroNeuralNetworks is an open-source Python project designed to introduce beginners to neural networks and demonstrate their application in stock price prediction. It guides users through the entire machine learning workflow, from data acquisition and preprocessing to model training and backtesting. The project includes implementations of Multilayer Perceptron (MLP) and Long Short-Term Memory (LSTM) models, explaining their relevance for time-series data like stock prices. While not intended for live trading, it serves as an educational template for understanding neural network fundamentals and can be extended for more sophisticated trading strategies. The project emphasizes the importance of data quality and provides a clear, step-by-step approach to building and evaluating predictive models.

ir-sim is an open-source, Python-based lightweight robot simulator specifically designed for navigation, control, and learning applications. It offers a simple and user-friendly framework that includes built-in collision detection, making it ideal for academic and educational use. The simulator allows for rapid prototyping of robotics and learning algorithms in custom scenarios with minimal coding and hardware requirements. Key features include the ability to simulate various robot platforms with diverse kinematics and sensors, quick scenario configuration using straightforward YAML files, and visualization of simulation outcomes with a naive visualizer for immediate debugging. It also supports multi-agent/robot learning projects.

Isaac Lab is a GPU-accelerated, open-source framework designed to unify and simplify robotics research workflows, including reinforcement learning, imitation learning, and motion planning. Built on NVIDIA Isaac Sim, it combines fast and accurate physics and sensor simulation, making it an ideal choice for sim-to-real transfer in robotics. The framework provides developers with essential features for accurate sensor simulation, such as RTX-based cameras, LIDAR, and contact sensors. Its GPU acceleration enables faster complex simulations and computations, crucial for iterative processes like reinforcement learning. Isaac Lab supports over 16 robot models and more than 30 ready-to-train environments, compatible with popular reinforcement learning frameworks like RSL RL, SKRL, RL Games, and Stable Baselines. It can run locally or be distributed across the cloud, offering flexibility for large-scale deployments.