Coding & Development

Erasing is an open-source project designed to remove specific concepts from diffusion models, offering a powerful way to fine-tune AI image generation. The tool provides updated code with diffusers support, significantly reducing GPU memory usage and increasing training speed by 5-8 times compared to older versions. It supports various diffusion models including SDv1.4, SDXL, FLUX, and FLUX.2 Klein, allowing users to erase entire concepts or precise attributes from concepts (e.g., removing hats from cowboys). The project includes installation guides, training instructions, and scripts for generating images and running a local Gradio demo, making it valuable for AI researchers and developers working with generative models.

OctNet is an open-source framework designed for deep learning with sparse 3D data, utilizing efficient space partitioning structures known as octrees. This approach significantly reduces the memory and compute requirements of 3D convolutional neural networks, allowing for the development of deep networks at high resolutions. By hierarchically partitioning space and storing pooled feature representations in leaf nodes, OctNet focuses memory allocation and computation on relevant dense regions. This enables deeper networks without sacrificing resolution, making it suitable for tasks such as 3D object classification, orientation estimation, and point cloud labeling. The framework includes core CPU and GPU code for network operations, data pre-processing tools, and a Torch wrapper for full network integration.

ObjectBox Java is a high-performance, lightweight on-device database and vector database specifically designed for Java and Kotlin applications. It enables efficient data storage and management within Android and JVM environments, including Linux, macOS, and Windows. A key differentiator is its first-on-device vector database capability, allowing for fast vector search to power RAG AI, generative AI, and similarity search applications. ObjectBox boasts exceptional speed, outperforming alternatives like SQLite and Realm in CRUD operations, while maintaining minimal CPU, power, and memory consumption. It simplifies development with a concise, language-native API that eliminates complex SQL queries, supports plain objects (POJOs) with built-in relations, and offers automatic schema migrations.

Emotion-recognition-neural-networks is an open-source project developed for emotion recognition using deep neural networks, specifically with TensorFlow. It employs convolutional neural networks (CNNs) for mood recognition, utilizing the FER-2013 Faces Database which contains 28,709 pictures across 7 emotional expressions. The project provides scripts for data transformation from CSV to NumPy, and supports training models using architectures like AlexNet. While the repository notes that the code might not be actively maintained or fully functional, it serves as a foundational academic project for those interested in exploring DNN-based emotion recognition.

e3nn is an open-source, modular framework designed to facilitate the development of neural networks with Euclidean symmetry. It provides fundamental mathematical operations such as tensor products and spherical harmonics, essential for building E(3) equivariant neural networks. The library is under active development, with breaking changes indicated by version number increments. It is recommended to install using pip, and users can contribute to its development or seek help through discussions and bug reports on GitHub. The framework is backed by research papers on Euclidean Neural Networks and e3nn itself, with BibTeX entries available for citation.

mPLUG-Owl is a family of multi-modal large language models (MLLMs) designed to enhance language models with multimodality through a modular approach. The project includes several iterations: mPLUG-Owl, mPLUG-Owl2, and mPLUG-Owl3, each building upon the previous version to offer improved capabilities. mPLUG-Owl2, for instance, was accepted by CVPR 2024 as a Highlight, and mPLUG-Owl2.1 provides a Chinese-enhanced version. The latest iteration, mPLUG-Owl3, focuses on long image-sequence understanding. The source code and weights for these models are available on HuggingFace, making them accessible for researchers and developers to integrate and experiment with.

mteb (Massive Text Embedding Benchmark) is an open-source Python library designed for comprehensive evaluation of text and multimodal embeddings. It offers a standardized framework to benchmark the performance of different embedding models across a wide array of tasks, including classification, clustering, semantic textual similarity (STS), retrieval, and reranking. The tool supports both monolingual and multilingual evaluations, with a focus on reproducibility and ease of use. Developers and researchers can use mteb to select models, define custom models, run evaluations, and analyze results, contributing to an interactive leaderboard that tracks the state-of-the-art in embedding performance. Its modular design allows for easy integration of new models, datasets, and benchmarks.

OllamaSharp offers .NET bindings for the Ollama API, making it straightforward to integrate Ollama into .NET applications for both local and remote interactions. It provides comprehensive API coverage, including chats, embeddings, model management (listing, pulling, pushing, copying, deleting, showing), and real-time streaming of responses and progress reports. The library is designed for ease of use, powering Microsoft Semantic Kernel and .NET Aspire, and supports advanced features like a sophisticated tools engine with source generators, multi-modality for vision models, and native AOT for improved performance. It also integrates seamlessly with Microsoft.Extensions.AI, allowing developers to use OllamaSharp as an IChatClient or IEmbeddingGenerator alongside other AI providers.

The Hugging Face Evaluation Guidebook is a comprehensive resource for understanding and implementing Large Language Model (LLM) evaluation. It provides both practical insights and theoretical knowledge, drawing from the experience of managing the Open LLM Leaderboard and designing the lighteval framework. The guidebook covers various evaluation methods, including automatic benchmarks, human evaluation, and LLM-as-a-judge approaches. It offers guidance on designing custom evaluations, troubleshooting common issues, and provides tips and tricks for both beginner and advanced users. Additionally, it includes sections on general LLM knowledge, such as model inference and tokenization, making it a valuable resource for anyone looking to ensure their LLM performs effectively.

DyNet is a powerful open-source neural network library, primarily developed by Carnegie Mellon University, with contributions from many others. Written in C++ and offering Python bindings, it's engineered for efficiency on both CPU and GPU architectures. A key differentiator is its ability to handle dynamic neural network structures, which can adapt and change for each training instance. This makes DyNet particularly well-suited for complex natural language processing tasks, where it has been successfully applied to build state-of-the-art systems for syntactic parsing, machine translation, and morphological inflection. The toolkit provides comprehensive documentation, tutorials for both C++ and Python, and examples to help users get started with its auto-batching feature and other functionalities.

opencode-openai-codex-auth is an open-source OAuth authentication plugin designed for developers seeking personal coding assistance through ChatGPT Plus/Pro subscriptions. This tool leverages OpenAI's official authentication method, ensuring secure and direct access to a wide range of Codex models. It simplifies the setup process, offering a "one config, every model" philosophy that makes accessing GPT-5.x and Codex models effortless. Key features include support for 22 model presets across GPT-5.2, GPT-5.2 Codex, and GPT-5.1 families, a variant system, and multimodal input capabilities. The plugin also provides usage-aware errors and automatic token refresh, enhancing the development workflow. It is specifically built for individual developers who prioritize simplicity and direct integration with their OpenAI subscriptions for coding tasks.

DropoutUncertaintyExps is an open-source project containing the experimental code for the paper "Dropout as a Bayesian Approximation: Representing Model Uncertainty in Deep Learning." The repository provides a framework for researchers to replicate and build upon the uncertainty experiments, with adaptations reflecting community feedback and bug fixes. It is based on José Miguel Hernández-Lobato's work on probabilistic backpropagation for scalable learning of Bayesian Neural Networks. The code utilizes datasets from the UCI machine learning repository, with specific data splits to ensure comparability of results. It details the methodology for hyperparameter tuning using grid-search and reports RMSE and log-likelihood metrics for various datasets, offering a valuable resource for academic research in deep learning uncertainty.

ego-planner-swarm is an open-source, efficient single/multi-agent trajectory planner specifically designed for multicopters. This tool extends the capabilities of EGO-Planner for swarm navigation, offering a fully autonomous and decentralized solution for multi-robot navigation in complex, unknown environments using only onboard resources. It supports ROS integration and is compatible with Ubuntu 16.04, 18.04, and 20.04, with a dedicated ROS2 version available on a separate branch. Developers can easily compile and run simulations, with options to configure for GPU usage for depth image generation or CPU for broader compatibility. The project also provides recommendations for optimizing CPU performance for stable computation times, making it a robust solution for advanced robotics development.

finetrainers is a work-in-progress library from Hugging Face designed for scalable and memory-optimized training of diffusion models. It provides support for various commonly used training algorithms, including DDP, FSDP-2, HSDP, and CP. Key features include LoRA and full-rank finetuning, conditional control training, and memory-efficient single-GPU training. The library also supports multiple attention backends like flash, flex, sage, and xformers, along with auto-detection of common dataset formats. It's built to handle combined image/video datasets, multi-resolution bucketing, and offers memory-efficient precomputation. finetrainers is recommended for use with PyTorch 2.5.1 or above for optimal performance and reproducibility.

ExLlamaV3 is an inference library specifically designed for running Large Language Models (LLMs) locally on modern consumer-class GPUs. Its headline feature is the new EXL3 quantization format, which is based on QTIP from Cornell RelaxML, allowing for efficient model conversion in a single step. The library supports flexible tensor-parallel and expert-parallel inference setups, and provides an OpenAI-compatible server via TabbyAPI for local or remote inference. It also includes features like continuous, dynamic batching, HF Transformers plugin support, speculative decoding, and 2-8 bit cache quantization. ExLlamaV3 aims to make advanced quantization techniques more accessible and less resource-intensive, enabling users to run large models like Llama-3.1-70B with minimal VRAM.

Model Madness is an experimental platform that pits leading AI models—Claude, ChatGPT, Gemini, and Grok—against each other in predicting the NCAA March Madness tournament. The experiment, conducted on the opening day of the tournament, assesses each AI's ability to generate an accurate bracket based on real-world data. It highlights significant differences in AI performance, with some models fabricating teams or misplacing them, requiring multiple correction prompts. The tool tracks live results, compares AI picks to public percentages, and provides a visual bracket view, offering insights into AI's current capabilities and limitations in handling dynamic, factual information.

Mind AI is a neuro-symbolic AI platform designed for complete logical reasoning, distinguishing itself from traditional AI through its controllability, explainability, and reasoning abilities. Every process within Mind AI is transparently verifiable, allowing for precise adjustments to AI models. As a cornerstone of Neuro-symbolic AI, it seamlessly integrates with conventional AI to enable Hybrid Intelligence. The platform's proprietary Canonical technology integrates abduction, deduction, and induction to construct advanced Wisdom Graphs, which convert unstructured natural language into structured logical flows. This approach embodies Hybrid Intelligence, combining Symbolic AI's accuracy with neural network scalability, and aims to make controllable and trustworthy AI accessible for everyone.

Fara-7B is Microsoft's first agentic small language model (SLM) specifically engineered for computer use. With only 7 billion parameters, it offers an ultra-compact solution for automating multi-step tasks on behalf of users. Unlike traditional chat models, Fara-7B interacts with computer interfaces visually, perceiving webpages and performing actions like scrolling, typing, and clicking directly on predicted coordinates without relying on accessibility trees. This design allows for efficient on-device deployment, reducing latency and enhancing privacy by keeping user data local. Fara-7B completes tasks efficiently, averaging only ~16 steps per task, and achieves state-of-the-art performance within its size class, competing with larger agentic systems. It is trained on 145K trajectories using a novel synthetic data generation pipeline built on the Magentic-One multi-agent framework, and is based on Qwen2.5-VL-7B with supervised fine-tuning.

Towhee Examples offers a diverse collection of applications designed to analyze unstructured data using the Towhee framework. These examples cover a wide range of tasks, such as reverse image search, reverse video search, audio classification, and question and answer systems. Additionally, it includes applications for molecular search and deepfake detection. The platform aims to democratize the process of generating embedding vectors (x2vec) by providing easily runnable examples that leverage machine learning models and operations. It supports various models like ResNet, VGG, EfficientNet, ViT for image tasks, DPR for NLP, and Pytorchvideo for video. This resource is ideal for developers and data scientists looking to implement advanced data analysis solutions.

Food-Recipe-CNN is a deep learning project designed to recognize food images and suggest matching recipes. Utilizing deep convolutional neural networks (CNNs) with Keras, this system can classify food images into 230 distinct categories. The project leverages a large dataset of over 400,000 food images and 300,000 recipes from chefkoch.de. It employs transfer learning with pre-trained CNNs like InceptionV3 and VGG16, alongside feature extraction and dimensionality reduction techniques such as PCA. The goal is to provide a solution for automated recognition of photographed dishes and subsequent recipe retrieval, with a web application called DeepChef in development.

Panicle Tech is a leading software development company offering specialized services in AI/ML solutions, Cloud infrastructure, DevOps, AWS services, and Security integration. They provide expert consultancy to help businesses build scalable startups and robust enterprise solutions. Their expertise spans various technologies including Kubernetes, Docker, TensorFlow, Python, and React, ensuring comprehensive support for modern software development needs. Panicle Tech focuses on delivering cutting-edge technology and strategic guidance to empower clients in navigating complex technological landscapes and achieving their business objectives.

FlexFlow Train is an open-source deep learning framework designed to accelerate distributed deep neural network (DNN) training. It achieves this by automatically searching for and implementing efficient parallelization strategies. The tool helps optimize the training process, reducing the time required for model development and improving overall efficiency. It supports various deep learning models and hardware configurations, making it a versatile solution for researchers and developers working with large-scale DNNs. The project is developed and maintained by teams from several prominent institutions, including CMU, Facebook, Los Alamos National Lab, MIT, Stanford, and UCSD.

open-llms is a comprehensive GitHub repository that serves as a curated list of open Large Language Models (LLMs) explicitly licensed for commercial use, including Apache 2.0, MIT, and OpenRAIL-M. This resource is invaluable for developers, researchers, and businesses looking to integrate open-source LLMs into their applications without licensing concerns. The repository details each model's release date, available checkpoints, associated research papers or blog posts, parameter sizes, context lengths, and specific licenses. It also includes a dedicated section for open LLMs tailored for code generation, offering insights into models like SantaCoder, CodeGen2, and StarCoder. Contributions to the list are welcomed, ensuring it remains up-to-date with the latest commercially viable open LLM releases.

Leia leverages artificial intelligence to empower users in rapidly building and deploying custom digital experiences and websites. This platform simplifies the process of creating tailored online content and customer interactions, making web development accessible to all skill levels. It focuses on streamlining the management of a business's online presence through intelligent automation. The tool aims to reduce the complexity and time involved in web development, allowing users to focus on content and strategy rather than intricate coding. By providing AI-powered assistance, Leia helps users create and manage their online presence efficiently and effectively.