Projects

Overview: This project is a design-fidelity QA tool for DeepSeek Harness, designed to evaluate whether text-based models match their mockups. It includes four fixtures and 23 injected defects to ensure the model's correctness. Key features: - Supports any text-based model that can call the `deepseek_vision` tool. - Reads and evaluates images uploaded by the user. - Provides a benchmark for the model's performance. - Retires itself when DeepSeek ships its vision. Tech stack: JavaScript Why it matters: This tool ensures that models like DeepSeek Harness can effectively perform product design tasks, as they lack visual capabilities. By integrating design fidelity QA, it helps developers ensure that the models' outputs are accurate and visually consistent with the design specifications. Repository: sunxin-ai/dsh-design-qa GitHub stars: 44

### Overview: Atlas is a system designed for continual learning from agent workflows, enabling a closed-loop between real-world agent execution and model improvement. The platform consists of two main components: the **Atlas SDK**, which wraps existing agent systems in a dual-agent reasoning loop, capturing causality data and streaming rich telemetry to a database; and the **Atlas Core**, an offline training engine that uses this data to train improved models via methods like on-policy distillation (GKD) and reinforcement learning (GRPO). ### Key features: - **Atlas SDK**: Wraps agent systems in a dual-agent reasoning loop, capturing causality data and streaming rich telemetry to a database. - **Atlas Core**: The offline training engine that uses the captured data to train improved models via methods like on-policy distillation (GKD) and reinforcement learning (GRPO). - **Training Methods**: Supports multiple training methods including GRPO, GKD, and SFT, each using the same Postgres backend. ### Tech stack: Python ### Why it matters: Atlas is a powerful tool for continuous improvement of artificial intelligence agents, enabling more efficient and effective learning from real-world interactions. ### Repository: Arc-Computer/ATLAS ### GitHub stars: 43 ### GitHub description: A Continual Learning Framework for Production LLM Agents

### Overview: Diagnose the performance of your RAG🩺 is a lightweight framework built with LlamaIndex that allows you to evaluate the performance of LLMs and retrieval models in RAG frameworks using your documents. It can be used as an application (thanks to FastAPI and Gradio) running locally on your machine, or as a Python package. ### Key features: - **Application (Docker):** A lightweight application that can be launched using Docker, providing a user-friendly interface for evaluating RAG performance. - **Python Package:** A Python package that can be installed using pip, offering flexibility and ease of integration into existing projects. - **Integration with Qdrant Client:** The framework integrates with Qdrant Client, a popular vector database, enabling efficient storage and retrieval of documents. ### Tech stack: - **Python:** The primary language used for development. - **LlamaIndex:** The core library used to build the framework. - **FastAPI:** A modern web framework for building APIs. - **Gradio:** A tool for building interactive web applications. - **Qdrant Client:** A vector database client for efficient storage and retrieval of documents. ### Why it matters: Diagnose the performance of your RAG🩺 is essential for developers and researchers who need to evaluate the effectiveness of their RAG frameworks. It provides a straightforward way to assess the performance of LLMs and retrieval models, ensuring that your models are optimized for better results.

### Overview: MultiNERD is a novel methodology for automatically creating multilingual, multi-genre, and fine-grained annotations for Named Entity Recognition and Entity Disambiguation. It is the first language-agnostic approach that covers 10 languages, 15 NER categories, and 2 textual genres. The dataset is designed to address the limitations of existing NER datasets, enabling the development of more versatile and accurate NLP systems. ### Key Features: - **10 Languages**: Chinese, Dutch, English, French, German, Italian, Polish, Portuguese, Russian, and Spanish. - **15 NER Categories**: Person (PER), Location (LOC), Organization (ORG), Animal (ANIM), Biological entity (BIO), Celestial Body (CEL), Disease (DIS), Event (EVE), Food (FOOD), Instrument (INST), Media (MEDIA), Plant (PLANT), Mythological entity. - **2 Textual Genres**: Fiction and Non-Fiction. ### Tech Stack: - **Data and Evaluation Code**: Available on Hugging Face Datasets. ### Why It Matters: MultiNERD represents a significant advancement in the field of NLP by providing a language-agnostic methodology for creating multilingual, multi-genre, and fine-grained annotations for Named Entity Recognition and Entity Disambiguation. This dataset is crucial for researchers and practitioners aiming to improve the accuracy and versatility of NLP models, particularly in the context of multilingual and multimodal applications.

Overview: LLM-I is a framework that transforms large language models (LLMs) into versatile creators capable of generating natural interleaved multimodal content, including images, text, and code. Key features: - Supports online image search for factual grounding - Utilizes diffusion-based generation for creative synthesis - Enables code execution for data visualization - Performs image editing on existing content Tech stack: Python Why it matters: LLM-I simplifies the process of creating multimodal content by integrating LLMs into a tool-use framework, making it easier to generate and manipulate images, text, and code seamlessly. Repository: ByteDance-BandAI/LLM-I GitHub stars: 43 GitHub description: 🚀 LLM-I: Transform LLMs into natural interleaved multimodal creators! ✨ Tool-use framework supporting image search, generation, code execution & editing

### Overview: This project presents a novel approach to language-driven 6-DoF grasp detection in cluttered point clouds, utilizing negative prompt guidance. The method aims to enhance the fine-grained challenge of detecting specific objects by steering away from undesired ones, thereby improving the accuracy and efficiency of the grasp detection process. ### Key features: - Novel diffusion model incorporating negative prompt guidance learning. - Improved language-driven grasp detection in cluttered environments. - Assists in fine-tuning the grasp detection process for desired objects. - Enhances the accuracy and efficiency of grasp detection tasks. ### Tech stack: - Python ### Why it matters: This project introduces a significant advancement in the field of computer vision and robotics, offering a more precise and efficient method for detecting specific grasps in complex environments. The use of negative prompt guidance not only improves the accuracy of grasp detection but also streamlines the process, making it more practical and effective for real-world applications.

### Overview: This repository introduces Scalable Group Inference, a method to generate diverse and high-quality sets of outputs using diffusion models. It addresses the issue of redundant results when sampling from generative models independently, offering a scalable solution for applications requiring multiple outputs. ### Key Features: - Scalable Group Inference - Efficient progressive pruning - Solves a quadratic integer programming (QIP) problem - Generates diverse and high-quality outputs ### Tech Stack: Python ### Why it Matters: This method allows for generating a set of outputs that are diverse and high-quality, addressing the issue of redundant results when using I.I.D. sampling. It provides a scalable solution for applications requiring multiple outputs, making it particularly useful in creative exploration and diverse image generation.

Overview: Marvin is a dynamic virtual assistant built with Python that allows users to easily add new voice commands without modifying the code. Key features: - Open websites and perform research - Play music (with an invisible player) - Set reminders Tech stack: Python Why it matters: Marvin simplifies adding new voice commands by updating the `commands.json` file, making it accessible for easy modification without needing to change the code. Repository: Gust4voSales/Marvin-VirtualAssistent GitHub stars: 43 GitHub description: A dynamic virtual assistant built with Python, featuring easy addition of voice commands without code changes.

Overview: A baseline model for multimodal classification that combines text and images using BERT and VGG16. Key features: - Text representation: BERT's average hidden vectors or the CLS token's vector. - Image representation: VGG16's 4096-dimensional feature vector. - Concatenated text and image features passed through a MultiModal Gated Layer. - Used datasets: Hateful memes detection from Facebook Challenge and Multimodal IMDb. Tech stack: Jupyter Notebook Why it matters: This model provides a simple yet effective baseline for multimodal classification tasks, demonstrating the potential of combining text and images for classification tasks. Repository: IsaacRodgz/ConcatBERT Primary language: Jupyter Notebook GitHub stars: 43

### Overview: DiffuLab is designed to simplify the training and customization of diffusion models, providing a flexible and easy-to-use framework. ### Key Features: - **Architecture**: Supports various architectures like U-Net and DiT. - **Formalization**: Offers different formalizations such as Rectified Flow and DDPM. - **Sampler for Inference**: Allows for flexible inference using different samplers. ### Tech Stack: Python ### Why it Matters: DiffuLab simplifies the process of training and customizing diffusion models, making it accessible and customizable for developers.

### Overview: ARC-Decode accelerates speculative decoding by extending acceptance under risk-bounded conditions, significantly improving efficiency on sampling tasks. ### Key Features: - **ARC-Decode**: A training-free speculative decoding framework that accelerates generation under sampling by extending acceptance under bounded risk. - **ICML 2026 Poster**: Official implementation of ARC-Decode, presented at ICML 2026. - **Python**: Primary language used for implementation. ### Tech Stack: - Python ### Why It Matters: ARC-Decode provides a practical solution for accelerating speculative decoding tasks, particularly under sampling conditions, by leveraging risk-bounded acceptance. This approach maintains efficiency while ensuring accurate results, making it valuable for applications requiring high throughput and low latency.

Overview: VulkanIlm is a Python library and CLI designed to run LLMs locally using Vulkan GPU acceleration, making it accessible to legacy and non-NVIDIA hardware. Key features: - Significant speedups on legacy GPUs and iGPUs - Broad GPU support: AMD, Intel, NVIDIA (via Vulkan) - Python-first API and easy CLI tools - Auto detection and GPU-specific optimizations - Auto build/install of `llama.cpp` Vulkan backend - Real-time streaming token generation - Reproducible benchmark scripts Tech stack: Python Why it matters: VulkanIlm provides a Pythonic way to run AI with knowledge (Ilm) on legacy Vulkan-capable systems, making it an essential tool for developers working with legacy hardware. Repository: Talnz007/VulkanIlm GitHub stars: 43 GitHub description: GPU-accelerated LLaMA inference wrapper for legacy Vulkan-capable systems a Pythonic way to run AI with knowledge (Ilm) on fire (Vulkan).

### Overview: LLaVA-Qwen2 is a project that enhances the Qwen2 base model with advanced capabilities, specifically tailored for visual instruction tuning. It leverages the Qwen2 model to achieve superior performance in large language and vision tasks. ### Key Features: - **Visual Instruction Tuning**: Enhances the Qwen2 model to better understand and respond to visual instructions. - **Fine-Tuning**: Integrates the Qwen2 model for improved performance in various language and vision tasks. - **Advanced Capabilities**: Leverages the capabilities of the Qwen2 model to achieve superior results in tasks such as art, business, science, health, and more. ### Tech Stack: - Python ### Why It Matters: LLaVA-Qwen2 is designed to improve the performance of large language and vision models by incorporating the advanced capabilities of the Qwen2 base model. This project is particularly useful for researchers and developers looking to enhance their models for specific tasks, such as visual instruction tuning and fine-tuning for improved performance in various domains. ### Repository: - **Repository**: TobyYang7/Llava_Qwen2 - **GitHub Stars**: 43 - **Primary Language**: Python ### Download: ```bash git lfs install git clone https://www.modelscope.cn/TobyYang7/llava-qwen2-1.5b-instruct-finvis.git ``` ### MMMU Eval: ```bash bash eval.sh ``` ### LLaVA-Qwen2-1.5B Result: | Subject |

Overview: This project introduces Autoregressive Denoising Diffusion Models (ADDM) for multivariate probabilistic time series forecasting, benchmarking datasets from various domains including medical, environmental, financial, and neural network forecasting competitions. Key features: - Medical: Hospital, Covid - Environmental: Solar-energy - Financial: fred_md (Macroeconomic Financial Variables) - Neural Network Forecasting Competition: nn5_daily_with_missing Tech stack: Jupyter Notebook Why it matters: This project provides a comprehensive framework for evaluating and benchmarking time series forecasting models across different domains, offering valuable insights for researchers and practitioners alike. Repository: absaw/DDM_Timeseries_Forecast Primary language: Jupyter Notebook GitHub stars: 43

Overview: DadAI is an AI assistant designed specifically for first-time dads, providing emotional support, practical tips, and insights into common challenges during pregnancy and early parenthood. Key features: - Emotional support during pregnancy and early parenthood - Sleep deprivation, relationship strain, identity loss - Dad mental health, bonding struggles, work-life guilt - Couple conflict after baby, breaking generational patterns - Practical tips from real fathers who've been there Tech stack: Python Why it matters: DadAI offers a personalized and practical approach to support new dads, addressing their specific needs and providing relevant advice from real experiences.

### Overview: SenseVoiceSmall is a Rust-based speech recognition library that leverages Candle for non-RKNN runtime and RKNN for Rockchip NPU runtime. It provides pure Rust ASR, VAD, and RKNN paths, making it suitable for applications requiring high performance on Rockchip NPU. ### Key Features: - Pure Rust ASR path: Candle + official `model.pt` (native PT loading). - VAD path: Candle + official `funasr/fsmn-vad` `model.pt` (auto-downloaded by hf-hub). - RKNN path: keep `rknpu` backend for Rockchip NPU. - No external ONNX Runtime (`ort`) library is required. ### Tech Stack: Rust ### Why It Matters: This library simplifies speech recognition and VAD processing, making it highly efficient and easy to integrate into Rust-based projects. Its compatibility with Rockchip NPU ensures optimal performance on specific hardware, making it ideal for embedded systems and applications requiring low latency. ### Repository: Repository: darkautism/sensevoice-rs GitHub Stars: 43 GitHub Description: A Rust-based, SenseVoiceSmall ### Installation: - Install `rknn.so`: ```bash sudo curl -L https://github.com/airockchip/rknn-toolkit2/raw/refs/heads/master/rknpu2/runtime/Linux/librknnrt.so -o /lib/librknnrt.so ``` - Add the `rknpu` feature gate to `Cargo.toml`.

### Overview: This repository contains codes and datasets for the ICASSP 2023 paper titled "Evaluating parameter-efficient transfer learning approaches on SURE benchmark for speech understanding". The repository introduces a new adapter called ConvAdapter, which outperforms standard adapters while showing comparable performance against prefix tuning and LoRA with only 0.94% of trainable parameters on some tasks in the SURE benchmark. ### Key Features: - **ConvAdapter**: A new adapter based on 1D convolution that significantly reduces the number of trainable parameters compared to standard adapters. - **SURE Benchmark**: A speech understanding benchmark specifically designed for parameter-efficient learning. - **Parameter-Efficient Transfer Learning**: Techniques such as adapters and prefix tuning are explored to improve the efficiency of transfer learning for speech processing tasks. ### Tech Stack: - Python ### Why It Matters: This work demonstrates the effectiveness of parameter-efficient transfer learning techniques in improving the efficiency and performance of speech processing tasks, particularly in reducing the computational cost and overfitting issues associated with fine-tuning large pre-trained models.

Overview: This repository is an ONNX-compatible Fast SeamlessM4T, designed for massively multilingual and multimodal machine translation using speech and text. It aims to facilitate effortless communication across different linguistic communities. Key features: - Text Encoder - Speech Encoder - Text Decoder - Unit Encoder - Unit Decoder - Vocoder Tech stack: Python Why it matters: This project provides a practical solution for multilingual and multimodal machine translation, making it easier for people from diverse linguistic backgrounds to communicate effectively through speech and text. GitHub stars: 43 Credits: If you use any ideas from the papers or code in this repo, please cite the SeamlessM4T paper.

### Overview: This project involves the classification of 11 types of audio clips using Mel-frequency cepstral coefficients (MFCCs) features and Long Short-Term Memory (LSTM) networks. The model is pretrained on the Speech Command Dataset with extensive data augmentation techniques, including shift, amplify, and superposition noise. The final model achieves an accuracy of 100.0% on a custom dataset, demonstrating strong performance in classifying audio clips. ### Key Features: - Pretrained model on Speech Command Dataset - Intensive data augmentation - Classification of 11 audio clip types - Random splits for training and testing: 0.7/0.3 and 0.9/0.1 ### Tech Stack: - Python - PyTorch - Jupyter Notebook ### Why it Matters: This project showcases the effectiveness of combining MFCCs features with LSTM networks for audio classification, demonstrating high accuracy and robustness against data augmentation techniques. It provides a practical example for implementing audio classification systems in machine learning projects.

### Overview: This collection features 252 curated Grok Imagine prompts across nine categories, each with preview images to help you see the results before running the prompts. ### Key Features: - **Portrait & Avatar**: 28 prompts - **Product & E-commerce**: 28 prompts - **Poster & Typography**: 28 prompts - **Infographic & Diagram**: 28 prompts - **Anime & Illustration**: 28 prompts - **3D & Material**: 28 prompts - **Food & Lifestyle**: 28 prompts - **Architecture & Interior**: 28 prompts - **Nature & Landscape**: 28 prompts ### Tech Stack: Unknown ### Why It Matters: This collection provides a curated set of Grok Imagine prompts tailored for various use cases, allowing developers to quickly generate visually appealing images without needing to write complex prompts themselves.