Projects

### Overview: Quiver is a pure Swift framework designed for statistics, linear algebra, and machine learning. It simplifies these complex mathematical concepts into a natural, Swift-friendly interface, making it accessible for both students and educators. Quiver runs on every Apple platform and is fully compatible with server-side Swift, ensuring seamless integration into various development environments. ### Key Features: - **Zero External Dependencies**: Quiver is completely self-contained, requiring no external libraries. - **Swift 5.9+**: Utilizes the latest Swift version for optimal performance and features. - **Educational Tool**: Provides a comprehensive set of models and algorithms for statistics, linear algebra, and machine learning, including real-world datasets. - **Model Exposability**: Models expose their learned parameters, allowing for easy visualization and understanding of the algorithms. - **Production Ready**: The framework is designed for production use, with models being `Codable` value types, ensuring compatibility across different platforms and devices. - **True Effort Score**: Tracks the learning process from start to finish, providing detailed insights into the model's performance and adjustments. ### Tech Stack: - **Swift**: The primary programming language for development. - **Platforms**: Supports iOS, macOS, watchOS, tvOS, and visionOS. ### Why It Matters: - **Ease of Use**: Simplifies complex mathematical concepts into a natural, Swift interface, making it easier for developers to learn and apply statistical and machine learning techniques. - **Educational Tool**: Ideal for students and educators, providing comprehensive models and datasets to enhance learning. - **Production Ready**:

Overview: pdfvision is a CLI, MCP server, and TypeScript library designed to extract text, layout, and page images from PDFs for AI agents. It provides detailed feedback on extraction failures, allowing agents to recover from silent failures and improve their reliability. Key features: - Extracts text, layout, and page images from PDFs - Provides detailed warnings on extraction failures - Offers recovery steps for failed extraction scenarios - Supports CLI, MCP server, and TypeScript library Tech stack: TypeScript Why it matters: pdfvision helps AI agents overcome silent failures in PDF extraction, ensuring reliable and human-like PDF vision. It surfaces text, layout, and page images together, making it easier for agents to recover from extraction errors and improve overall reliability. Repository: yamadashy/pdfvision GitHub stars: 24 GitHub description: CLI, MCP server, and TypeScript library that extracts text, layout, and page images from PDFs for AI agents. When extraction goes wrong, it says so per page and says what to do next.

Overview: LongBench-T2I is a comprehensive benchmark and agent framework designed to evaluate and improve complex instruction-based text-to-image generation, aiming to push the field towards AGI-level capabilities in controllable visual synthesis. Key features: - Comprehensive benchmark and agent framework for complex instruction-based image generation - Holistic evaluation and analysis of complex instruction-based image generation models - Structured summary of key insights for further research and development Tech stack: Python Why it matters: LongBench-T2I provides a robust platform for researchers and developers to test, evaluate, and improve complex instruction-based image generation models, ultimately contributing to the advancement of AGI-level capabilities in visual synthesis. Repository: yczhou001/LongBench-T2I GitHub stars: 25

Overview: This project, developed by yuruofeifei, is an MXNet Model Serving solution designed to simplify the deployment and management of machine learning models. The system allows for the installation, configuration, and serving of models using Python, making it accessible and easy to integrate into existing projects. Key features: - MXNet Model Serving integration - Support for multiple models - Ability to load input modules with custom preprocess and postprocess - Generation of OpenAPI client SDKs - Exporting existing models to serving model format Tech stack: Python Why it matters: This project streamlines the process of deploying machine learning models, making it easier for developers to integrate and manage models without needing deep knowledge of model deployment or server setup. It simplifies the management of models by providing a user-friendly interface and automated tools, thus enhancing the efficiency and accessibility of model deployment.

Overview: 本仓库提供加速框架的最简单实操，旨在帮助用户快速体验主流LLM加速框架的性能。 Key features: - UI界面：提供简单的gradio web界面，方便用户进行加速任务的提交和管理。 - openai接口：方便用户通过代码进行加速任务的提交和管理。 - 压力测试：提供压力测试脚本，方便用户进行加速任务的压力测试。 Tech stack: Python Why it matters: 本仓库简化了大模型推理框架的开发和使用，使得用户能够快速上手并体验到高性能的加速效果，从而加速了大模型的本地部署和应用。

### Overview: This repository contains a collection of papers that explore the implicit reasoning capabilities of Large Language Models (LLMs). These papers aim to enhance the ability of LLMs to reason without explicitly stating the reasoning steps, thereby improving efficiency and effectiveness. ### Key Features: - **Efficient Reasoning with Hidden Thinking** (2025) - **Scaling up Test-Time Compute with Latent Reasoning** (2025) - **CODI: Compressing Chain-of-Thought into Continuous Space via Self-Distillation** (2025) - **SoftCoT: Soft Chain-of-Thought for Efficient Reasoning with LLMs** (2025) - **Scalable Language Models with Posterior Inference of Latent Thought Vectors** (2025) - **Reasoning with Latent Thoughts: On the Power of Looped Transformers** (2025) - **CoT-Valve: Length-Compressible Chain-of-Thought Tuning** (2025) - **LightThinker: Thinking Step-by-Step Compression** (2025) - **Beyond Words: A Latent Memory Approach to Internal Reasoning in LLMs** (2025) - **From Explicit CoT to Implicit CoT: Learning to Internalize CoT Step by Step** (2024) - **Training Large Language Models to Reason in a Continuous Latent Space** (2024) - **Compressed Chain of Thought: Efficient Reasoning Through

### Overview: CyberBrain_Model is an advanced AI project designed for fine-tuning the model `DeepSeek-R1-Distill-Qwen-14B` specifically for cyber security tasks. This repository provides tools and scripts for training and fine-tuning large language models efficiently using minimal hardware resources. ### Key Features: - Fine-tuning of `DeepSeek-R1-Distill-Qwen-14B` for cyber security applications - Efficient training on devices with limited computational power - Accessible training data from various cyber security sources - Documentation on mapping and table ways - Scripts for loading the model, configuring training arguments, and starting the training process ### Tech Stack: Jupyter Notebook ### Why it Matters: CyberBrain_Model aims to adapt the model for ethical cyber security applications, making it efficient even on devices with limited computational power, whether you have a low-end CPU or a GPU with limited VRAM. Repository: 0PeterAdel/CyberBrain_Model GitHub Stars: 24

### Overview: Local-LLM 是一个支持 ChatGLM2 和 Llama-2 的本地模型服务，旨在为开发者提供便捷的模型部署与启动体验。 ### Key features: - 支持 ChatGLM2 和 Llama-2 模型 - 一键安装与启动，无需额外配置 - 提供 WebUI 和 API 模式供选择 - 支持多种模型下载与配置 - 易于安装与部署，适用于各种操作系统 ### Tech stack: Python ### Why it matters: Local-LLM 提供了一个简单易用的解决方案，让开发者能够轻松部署和使用大型语言模型，无需依赖云服务。这在开发过程中节省了大量的时间和成本，同时提供了高度的灵活性和可定制性。

### Overview: This repository provides a comprehensive guide and practical approach to using reinforcement learning (RL) techniques with large language models (LLMs) to improve alignment and performance. It includes detailed explanations of the post-training process, reinforcement learning methods, and how to create and train your own RL environments using Jupyter Notebooks. ### Key Features: - **Reinforcement Learning with Verifiable Rewards (RLHF)**: Techniques for optimizing language models to better align with human preferences and reduce bias. - **Post-Training Alignment**: Methods for refining language models post-training to improve their alignment with human values. - **Jupyter Notebook Implementation**: Practical examples and code snippets to demonstrate how to implement these techniques in a Jupyter Notebook environment. ### Tech Stack: Jupyter Notebook ### Why It Matters: Reinforcement learning is crucial in post-training alignment of language models to ensure they better reflect human values and preferences. This guide provides a practical approach to implementing these techniques, making it valuable for developers and researchers working on LLMs and alignment. ### Repository: **Repository:** ALucek/rl-for-llms **GitHub Stars:** 24 **GitHub Description:** Context & Guide For Reinforcement Learning with Verifiable Rewards with Large Language Models **Primary Language:** Jupyter Notebook ### Training Sequence: 1. **Pre-Training**: Large language model pre-training using next-token prediction. 2. **SFT**: Base instruction-following + tools + formatting. 3. **Preference Training**: Push general answer quality and safety. 4. **Targeted RL**: Optimization for specific tasks

### Overview: OpenCLAW SEED is an autonomous self-evolving research agent that continuously improves through LoRA fine-tuning on harvested research papers. This project is part of the P2PCLAW ecosystem and uses the Apache 2.0 license. ### Key features: - Continuous LoRA fine-tuning on research papers - Self-evolving from SmolLM2-135M to Qwen2.5-7B+ - Trains on 10,000+ research papers - Uses the Hugging Face Transformers library - Part of the P2PCLAW ecosystem ### Tech stack: Python ### Why it matters: This project represents an innovative approach to continuous learning and research automation, leveraging the power of pretrained models and fine-tuning on real-world data to improve over time.

Overview: This project aims to familiarize developers with the ecosystem of large language models (LLMs) and quantization techniques, with a focus on cloud quantum computing. It provides a set of Jupyter notebooks that have been executed on the Kaggle platform using GPUs, covering various ML tasks and fine-tuning metrics. Key features: - Fine-tuning of large language models (LLMs) - Familiarization with LLMs tools ecosystem and quantization techniques - Coverage of ML tasks on the Kaggle platform - Metrics for fine-tuning process - Tools used for LLMs and quantization techniques Tech stack: Jupyter Notebook Why it matters: This project serves as a practical guide for developers interested in using deep neural networks on consumer-grade hardware, particularly focusing on researching GenAI. It provides insights into cloud quantum computing and the quantization techniques necessary for efficient LLMs operations. Repository: Aisuko/notebooks GitHub stars: 24 GitHub description: Implementation for the different ML tasks on Kaggle platform with GPUs. README excerpt: # Overview > We might agree that the operation of LLMs will embed in daily programming in the future. So, we use these notebooks to familiarize ourselves with the LLMs tools ecosystem and quantization techniques. I believe that Cloud quantum computing is needed for the future of LLMs. Maybe somthing Qubernetes. All these notebooks have been completed running on the [Kaggle](https://www.kaggle.com/aisuko/code) platform. With the free GPUs. Some of

### Overview: DocPOI is a local chatbot designed to help manage and retrieve documents efficiently. It leverages your NVIDIA GPU for enhanced performance and security, ensuring absolute privacy and accessibility. ### Key Features: - **Document Retrieval:** Quickly find and access your documents. - **Auto Metadata Extraction:** Automatically extract and create metadata for your documents. - **Auto Naming:** Automatically generate appropriate names for your documents. - **PDF to PDF OCR:** Convert PDFs to searchable PDFs using OCR. - **Speech Recognition:** Interact with the assistant using voice commands. - **Voice Over (TTS):** Have documents and responses read aloud. ### Tech Stack: Python ### Why It Matters: DocPOI aims to simplify the process of managing and retrieving documents, making it a valuable tool for anyone dealing with a large volume of personal or professional files.

### Overview: DistRL: An Asynchronous Distributed Reinforcement Learning Framework for On-Device Control Agents This project introduces **DistRL**, a framework designed for efficient reinforcement learning in distributed environments. It supports flexible integration of agents and models, scalable data collection tools, and efficient online training with multi-machine support. ### Key Features: - **Framework Capabilities**: Supports flexible integration of agents and models, scalable data collection tools, and efficient online training with multi-machine environment support. - **Methodology Highlights**: - **DistRL**: An efficient reinforcement learning algorithm tailored for distributed environments. - **Baseline**: **DigiRL** (features automatic curriculum learning with doubly robust estimator filtering). - **Baseline**: **Filtered Behavior Cloning** (employs reward-based filtering). - **Agent Support**: Comprehensive support for both training and evaluation phases. - **Android-in-the-Wild Task Sets**: Tasks involving general browsing and app launching (AitW General), centered around shopping on popular e-commerce websites (AitW Web Shopping), and additional evaluations on other subsets. - **DDP Multi-GPU Training Support**: Facilitates multi-GPU training through `accelerate`, requiring only 15GB of GPU memory. ### Tech Stack: Python ### Why it Matters: This framework is crucial for enabling efficient reinforcement learning in distributed environments, particularly in scenarios where agents need to interact with complex, dynamic environments. It simplifies the process of training and evaluating agents in real-world applications, making it highly valuable for researchers and practitioners in the field of artificial intelligence and reinforcement learning.

Overview: MergeMix is a unified augmentation paradigm designed for visual and multi-modal understanding, presented at ICLR 2026. It offers a comprehensive solution for enhancing model performance across various tasks. Key features: - Unified augmentation techniques across visual and multi-modal tasks - Efficient implementation compatible with popular frameworks like PyTorch - Integrates seamlessly with existing models and frameworks - Supports both image classification and other multi-modal applications Tech stack: Python Why it matters: MergeMix provides a unified approach to data augmentation, enhancing model robustness and performance across different tasks, making it a valuable tool for researchers and practitioners in the field of computer vision and multi-modal understanding.

Overview: Attention-MoA is an official code for enhancing Mixture-of-Agents through semantic attention and deep residual synthesis, based on state-of-the-art models. Key features: - Enhances Mixture-of-Agents with semantic attention and deep residual synthesis - Supports large-scale and small-scale configurations - Evaluates on AlpacaEval 2.0, MT-Bench, and FLASK benchmarks Tech stack: Python Why it matters: Attention-MoA provides a framework for improving the performance of Mixture-of-Agents through advanced attention mechanisms and deep residual synthesis, making it a valuable tool for researchers and developers in the field of artificial intelligence and natural language processing.

### Overview: Comfy Cozy is a local counterpart to Comfy Cloud MCP, designed for VFX artists. It allows for zero-LLM recipes, surgical edits, and full undo capabilities, making it a key tool for artists looking to enhance their workflow. ### Key Features: - **Local Execution**: Renders on your GPU, against your models, on your disk. - **Surgical, Validated Edits**: Small, reversible changes that can be undone. - **Zero-LLM Recipes**: Instant, deterministic macros for "dreamier," "sharper," and "faster" effects. - **Provenance**: `workflow.lock` pins model SHA-256s and pack commits; drift warns during validate. - **6 Swappable LLM Brains**: Claude, GPT-4o, Gemini, Ollama, NVIDIA Nemotron, or any OpenAI-compatible endpoint. ### Tech Stack: Python ### Why It Matters: Comfy Cozy offers a local, production-grade alternative to Comfy Cloud MCP, providing artists with the flexibility to run their own rendering workflows without relying on remote servers. This makes it ideal for artists who want to maintain control over their rendering processes and can benefit from the speed and reliability of local execution. ### Repository: JosephOIbrahim/Comfy-Cozy ### GitHub Stars: 24 ### GitHub Description: Patent Pending & MIT License ### Why it Matters: Comfy Cozy is a key tool for VFX artists looking to enhance their workflow by providing zero-LLM recipes, surgical edits,

### Overview: SelfJev is a powerful tool that converts text and questions into structured decisions suitable for your application. It leverages a 4B model running on your own infrastructure, providing typed answers and probabilities from forward passes without generating written responses. This makes it ideal for integrating AI into your code without the need for a cloud service. ### Key features: - **Decision Engine**: Converts text into structured decisions, suitable for your application. - **Typed Answers**: Provides yes/no, choice, score, and multi-answer types. - **Probabilities**: Offers probabilities for each answer type. - **API Compatibility**: Compatible with Python SDK and Jev-compatible HTTP API. ### Tech stack: Python SelfJev is built using Python, leveraging the power of the Qwen3.5-4B model and LoRA for advanced decision-making capabilities. ### Why it matters: SelfJev simplifies integrating AI into your applications by providing a straightforward way to handle decisions without the need for a cloud service. Its self-hosted nature ensures you have full control over the infrastructure, making it ideal for environments where cloud services are not an option.

Overview: BuildView 是一个基于 Milvus、PostgreSQL、Redis、Neo4j 和大模型的私有化 RAG 问答系统，旨在为建筑工程提供知识库和项目智能解决方案。 Key features: - 支持智能问答、Agent 智能分析、知识图谱、施工图处理 - 使用 Docker 和 Python 进行容器化部署 - 基于 Milvus 的向量检索和混合检索 - 支持 OpenAI、Qwen、GLM、Ollama 等大模型接口 - 使用 Redis 和 PostgreSQL 进行缓存和日志管理 Tech stack: Python, Docker, Milvus, PostgreSQL, Neo4j, Redis, BGE, text2vec, OpenAI, OpenAI API, OpenAI LLM Why it matters: BuildView 提供了全面的建筑工程智能解决方案，通过结合 Milvus、PostgreSQL、Neo4j 和大模型，为用户提供高效、准确的问答和分析服务，从而提升项目的管理和决策效率。

### Overview: This implementation fine-tunes the Flan T5 large language model on the IMDB dataset for text classification, achieving an impressive accuracy of 93%. ### Key features: - Fine-tuned Flan T5 model - Text classification task on the IMDB dataset - Obtained 93% accuracy on evaluation set ### Tech stack: Transformers 4.28.1, Pytorch 2.0.0+cu118, Datasets 2.12.0, Tokenizers 0.13.3 ### Why it matters: This model demonstrates the effectiveness of fine-tuning large language models for practical applications, showcasing high performance in text classification tasks.

Overview: This project is an Enterprise-grade Knowledge Distillation framework designed for Persian Mental Health analysis. It distills reasoning capabilities from Llama-3-70b into a local Gemma-3-1b model using QLoRA and synthetic Chain-of-Thought datasets. Key features: - Knowledge Distillation Pipeline - Teacher-Student Architecture - Distillation via Synthetic Data - Rationale Distillation & Supervised Fine-Tuning (SFT) with QLoRA Tech stack: Python Why it matters: This project provides a specialized, lightweight AI model for detecting "Depression" vs. "Suicide" risk in Persian text, ensuring high diagnostic accuracy while maintaining high performance on consumer hardware. Repository: MDVR9980/Knowledge_Distillation_Project GitHub stars: 24