Projects

### Overview: Corus is a Python library designed to load and parse publicly available Russian corpora. It includes datasets such as the Lenta.Ru News Dataset, Lib.rus.ec, and Rossiya Segodnya, providing a comprehensive toolset for developers working with Russian language data. ### Key Features: - **Corus Library**: A Python library for loading and parsing Russian corpora. - **Datasets Included**: Lenta.Ru News Dataset (v1.0 and v1.1+), Lib.rus.ec, and Rossiya Segodnya. - **Usage Example**: The library can be used to load and iterate over news records from Lenta.Ru, with the option to specify the dataset version. ### Tech Stack: - Python 3.5+ and PyPy 3 ### Why It Matters: Corus simplifies the process of working with Russian language data by providing a standardized interface for accessing and processing publicly available datasets. This makes it easier for developers to integrate Russian language data into their projects without needing to manually download and parse large datasets.

### Overview: AI-based Protein Design has revolutionized the field, enabling accurate predictions of protein structures from amino acid sequences. This technology has the potential to significantly impact pharmaceutical research and drug development. ### Key Features: - **AlphaFold 2**: A tool for predicting protein structures from amino acid sequences, achieving atomic-level accuracy. - **Protein Structure Completion**: Methods that predict the structure of a protein given a partial structure. - **Fitted Energy Function with MD Simulation**: Combining AI models with molecular dynamics simulations for protein design. - **High-Level AI Papers**: A collection of research papers focusing on AI-based protein design, including those from prestigious journals like Nature, Science, and NeurIPS. ### Tech Stack: The repository tracks the latest developments in AI-based protein design, including tools like AlphaFold 2, protein structure prediction models, and AI algorithms for design. ### Why it Matters: This collection of papers highlights the potential of AI to revolutionize protein design, enabling more accurate predictions and potentially leading to new breakthroughs in pharmaceutical research and drug development.

Overview: This project is a chatbot web app and HTTP/WebSocket endpoints for LLM inference using the Petals client, designed for developers. Key features: - WebSocket API for real-time communication - HTTP API for batch requests - Petals client integration - Support for Llama 2 and other models Tech stack: Python Why it matters: This project provides a convenient way to integrate LLM inference with the Petals client, making it easier for developers to create chatbots and other applications that can leverage large language models. Repository: petals-infra/chat.petals.dev Primary language: Python GitHub stars: 319 GitHub description: 💬 Chatbot web app + HTTP and WebSocket endpoints for LLM inference with the Petals client

Overview: Ovis-Image is a 7B text-to-image model designed specifically for high-quality text rendering, optimized for efficient operation under stringent computational constraints. It excels in rendering text with precise alignment and correct spelling, making it suitable for applications like posters, banners, and UI mockups. Key features: - High-quality text rendering with precise alignment and correct spelling - Operates efficiently on a compact 7B parameter budget - Supports text-heavy, layout-sensitive prompts Tech stack: Python Why it matters: Ovis-Image is a powerful tool for generating high-fidelity text images, making it ideal for applications requiring precise typography and legibility, such as marketing materials, design, and UI development. Repository: ATH-MaaS/Ovis-Image Primary language: Python GitHub stars: 318 GitHub description: Ovis-Image is a 7B text-to-image model specifically optimized for high-quality text rendering, designed to operate efficiently under stringent computational constraints. README excerpt: <!-- <p align="center"> <img src=https://cdn-uploads.huggingface.co/production/uploads/637aebed7ce76c3b834cea37/3IK823BZ8w-mz_QfeYkDn.png width="30%"/> </p> --> <p align="center"> <img src="docs/imgs/ovis_image_title.png" width="40%"> </p> <!-- <h1 align="center"> Ovis-Image </h1> --> <p

Overview: This AI Smart Speaker leverages speech recognition, text-to-speech, and speech-to-text functionalities to enable voice and vision-driven conversations. It integrates with OpenAI and Langchain agents for additional web search capabilities. Key features: - Speech recognition for inputting prompts - Text-to-speech conversion of responses - Speech-to-text conversion for command recognition - Integration with OpenAI for advanced AI capabilities - Compatibility with Raspberry Pi and Python Tech stack: Python Why it matters: This speaker enhances user interaction by providing a seamless voice and text-based experience, making it ideal for personal assistants, educational tools, and more. Repository: Olney1/ChatGPT-OpenAI-Smart-Speaker GitHub stars: 318

### Overview: This repository contains all the tools and code used to build the ROOTS dataset, sourced and cleaned by the BigScience initiative to train the BLOOM models and a reduced version for tokenizer training. ### Key features: - Code for making the Pseudo-Crawl dataset - Filtering library for OSCAR - Code for preprocessing pipeline on crowdsourced dataset - Code for tokenizer's training dataset - Code for analysis and plots for the paper ### Tech stack: Jupyter Notebook ### Why it matters: This repository provides essential tools and code for the creation and preprocessing of the ROOTS dataset, which is crucial for training the BLOOM models and developing a tokenizer.

Overview: 本项目是一个自然语言处理项目，旨在通过对文本进行分类来实现对电商评论和司法罪名数据的自动分类。 Key features: - 文本预处理：对原始文本数据进行分词、转编码和长度统一等预处理步骤。 - 模型训练及预测：整合了预处理和神经网络的训练，支持模型的训练和预测功能。 Tech stack: Python Why it matters: 该项目展示了如何使用Python进行自然语言处理任务，包括文本预处理和分类模型的训练与预测，适用于电商评论和司法罪名数据的自动分类。

### Overview: Intelligent LLM-powered git commit message generator that understands your code! Supports multiple languages! ### Key features: - **35+ Supported Providers**: Anthropic, Azure OpenAI, Cerebras, ChatGPT (OAuth), Chutes.ai, Claude Code (OAuth), Crof.ai, DeepInfra, DeepSeek, Fireworks, Gemini, GitHub Copilot (OAuth), Groq, Kimi for Coding, LM Studio, MiniMax.io, Mistral AI, Moonshot AI, Neuralwatt, Ollama, Ollama Cloud, OpenAI, OpenCode Go, OpenRouter, Plexus Gateway, Qwen Cloud (CN & INTL), Replicate, Streamlake/Vanchin, Synthetic.new, Together AI, Wafer.ai, Z.AI (API & Coding Plans), Custom Endpoints (Anthropic/OpenAI). - **Smart LLM Analysis**: Understands intent, semantic awareness, intelligent filtering, and intelligent commit grouping. - **Multiple Message Formats**: One-liner, Standard, Verbose, 50/72 rule, DCO/Signoff. ### Tech stack: Python - Primary language: Python ### Why it matters: Automate your commits! Replace `git commit -m "..."` with `uvx gac` for contextual, well-formatted commit messages generated by large language models!

### Overview: AI Video Tutorial Generator is a powerful tool for creating and editing AI-driven video tutorials. It allows users to turn a question, topic, or source material into an editable lesson, providing a creative and interactive experience. ### Key Features: - **AI-Powered Presenters:** Generates realistic and anime-style presenters. - **Native Timeline Editor:** Allows users to refine the tutorial's timeline and quality. - **Local and Cloud AI Options:** Supports both local AI models and cloud-based AI. - **User-Friendly Interface:** Guides users through the planning and editing process, including audience selection, script creation, and media generation. ### Tech Stack: Python ### Why It Matters: AI Video Tutorial Generator streamlines the process of creating educational videos, making it easier for users to turn ideas into engaging tutorials. The tool supports a wide range of AI models and provides detailed feedback, ensuring high-quality output. This makes it an invaluable resource for educators and creators looking to enhance their teaching materials with AI-driven content.

### Overview: BayLing is an advanced multilingual language model designed to enhance language alignment and instruction following capabilities. It achieves superior performance in English/Chinese generation, instruction following, and multi-turn interactions. The model is based on the LLaMA architecture and has been trained with high-resource language instructions and cross-lingual instructions for over 100 languages, making it highly versatile and effective across various tasks. ### Key Features: - **Language Alignment**: BayLing enhances language alignment to ensure accurate and coherent responses across languages. - **Instruction Following**: It improves the ability to follow and execute instructions effectively, particularly in complex scenarios. - **Multi-turn Interaction**: BayLing excels in handling multiple turn interactions, providing seamless and contextually appropriate responses. ### Tech Stack: - **Python** - The primary programming language used for developing and training the model. ### Why It Matters: BayLing is pivotal in bridging cross-lingual alignment and instruction following, making it a valuable tool for multilingual applications. Its superior performance across various languages and tasks makes it a significant advancement in the field of large language models, particularly in the realm of English/Chinese language pairs. ### Repository: - **Repository**: BayLing-Models/BayLing - **GitHub Stars**: 317 - **GitHub Description**: "百聆" is an English/Chinese LLM equipped with advanced language alignment, showing superior capability in English/Chinese generation, instruction following, and multi-turn interaction. This project has been moved to: https://github.com/BayLing

Overview: BlenderLLM is a large language model designed to generate CAD scripts for Blender, enabling the creation of 3D models based on user instructions. The dataset includes 12,000 samples, with 2,000 manually annotated and 10,000 generated using GPT-4o. Key features: - BlenderLLM: Fine-tuned model for generating CAD scripts. - BlenderLLM Dataset: High-quality dataset of 12,000 samples. - CADBench: Benchmark for evaluating CAD-oriented LLMs. Tech stack: Python Why it matters: BlenderLLM provides a user-friendly interface and simplified input methods, making CAD-oriented LLMs accessible to a wider audience. It addresses the challenges of complex input forms in CAD applications and ensures high-quality, domain-specific datasets for training LLMs. The comprehensive evaluation framework and open-source nature of BlenderLLM contribute to the democratization of advanced CAD tools and secure, localized deployments.

### Overview: Delphic is a framework designed to simplify the integration and deployment of large language models (LLMs) using LlamaIndex and Langchain. It aims to provide a robust and scalable solution for querying and analyzing text data from documents. The project is MIT licensed and built with Python, leveraging LlamaIndex for indexing and Langchain for various functionalities. ### Key Features: - **Integration with LlamaIndex**: Facilitates seamless integration with LlamaIndex for efficient text querying. - **Langchain Compatibility**: Compatible with Langchain for additional functionalities and flexibility. - **OpenAI API Support**: Initially based on OpenAI's API, but future support for other LLMs is planned. - **Django Integration**: Built with Django for backend development, ensuring robust and scalable applications. - **PostgreSQL Vector Store**: Utilizes PostgreSQL as the vector store for efficient text indexing. ### Tech Stack: - Python - LlamaIndex - Langchain - OpenAI API - Django - PostgreSQL ### Why It Matters: Delphic provides a straightforward and efficient way to integrate large language models into applications, making it easier to analyze and manipulate text data from documents. The framework's compatibility with Langchain and support for various LLMs make it versatile and adaptable to different use cases. The integration with Django ensures a robust and scalable backend, while PostgreSQL serves as the vector store for efficient text indexing.

### Overview: picoLLM Inference Engine is a highly accurate and cross-platform SDK optimized for running compressed large language models. It significantly improves GPTQ by [significant margins](https://picovoice.ai/blog/picollm-towards-optimal-llm-quantization/) and is designed for on-device inference, making it ideal for applications requiring privacy and low latency. ### Key Features: - **Accuracy**: Highly accurate, improving GPTQ by [significant margins](https://picovoice.ai/blog/picollm-towards-optimal-llm-quantization/). - **Private**: LLM inference runs 100% locally, ensuring privacy and security. - **Cross-Platform**: Runs on Linux (x86_64), macOS (arm64, x86_64), and Windows (x86_64, arm64), as well as Raspberry Pi (3, 4, 5), Android, iOS, and Chrome, Safari, Edge, and Firefox. - **Runs on CPU and GPU**: Efficiently utilizes both CPU and GPU for optimal performance. - **Free for open-weight models**: No cost for open-weight models, making it accessible to a wide range of developers. ### Tech Stack: Python ### Why it Matters: picoLLM Inference Engine is a game-changer for developers looking to integrate large language models into their applications without compromising on accuracy, privacy, or performance. Its cross-platform compatibility and low latency make it an ideal solution for a variety of

### Overview: This project automates the overlaying of pitch motion and trajectory with machine learning, providing a real-time solution for baseball pitchers. ### Key features: - Automatically detects release points and traces trajectories. - Uses a fine-tuned YOLOv4 model for ball location. - Implements SORT tracking algorithm for individual ball tracking. - Applies image registration techniques to handle slight camera shifts. ### Tech stack: Python ### Why it matters: This project offers a practical solution for baseball pitchers by automating the overlaying process, making it easier to analyze and improve pitching techniques in real-time.

### Overview: LlamaDuo is an innovative solution for seamless migration of knowledge from service-oriented large language models (LLMs) to smaller, locally manageable models. This project, accepted to ACL 2025, showcases the effectiveness and scalability of LlamaDuo in various downstream tasks. ### Key Features: - **Seamless Migration**: Converts service-oriented LLMs into small, locally manageable models. - **Fine-Tuning**: Fine-tunes a small language model using a synthetic dataset generated by the service LLM. - **Multi-Turn Process**: Automatically improves the fine-tuned model using additional similar data generated by the service LLM. - **Practical and Scalable**: Ensures that the smaller model can match or surpass the service LLM's capabilities in specific tasks. ### Tech Stack: - **Python** - **Transformers** - **Peft** - **Alignment-Handbook** - **HuggingFace Hub** ### Why It Matters: LlamaDuo offers a practical and scalable solution for managing AI deployments in constrained environments, ensuring service continuity and privacy. By fine-tuning a small language model against a service-oriented LLM, LlamaDuo ensures that small models can match or surpass the capabilities of service LLMs in specific tasks, making it a valuable tool for AI deployment in various scenarios.

### Overview: This project explores the potential of large language models (LLMs) to assist in the training of clinical calculators using patient history, with a focus on the Wells' Criteria. The team aims to leverage open-source models to improve medical education and patient care. ### Key features: - **LLM Training for Clinical Calculators**: Utilizing LLMs to train clinical calculators based on patient history. - **Open-Source Models**: Leveraging open-source models to ensure reproducibility and accessibility. - **Wells' Criteria Integration**: Starting with the Wells' Criteria to demonstrate the application of LLMs in clinical decision-making. ### Tech stack: Python ### Why it matters: This project aims to bridge the gap between clinical practice and technological advancements, potentially revolutionizing medical education and patient care by leveraging the power of LLMs. By integrating open-source models and focusing on the Wells' Criteria, the project demonstrates the practical application of LLMs in clinical decision-making, making it a valuable contribution to the field.

Overview: This project is a BERT distillation experiment using both Keras and PyTorch, focusing on text classification tasks. The initial results show that the distilled model achieves an accuracy of 0.87 to 0.88 on the sentiment classification task for a clothing dataset, closely aligning with the expectations from the referenced paper. Key features: - BERT distillation - Text classification on a clothing dataset - Comparison with Keras and PyTorch implementations Tech stack: Python Why it matters: This project serves as a practical demonstration of BERT distillation, showcasing its effectiveness in reducing complexity while maintaining performance, which is crucial for practical applications in natural language processing tasks. Repository: qiangsiwei/bert_distill Primary language: Python GitHub stars: 316 GitHub description: BERT distillation（基于BERT的蒸馏实验 ） README excerpt:

Overview: This Python library provides a simple and efficient way to define GPU functions and run them in the cloud using the Runpod API and serverless worker SDK. It allows developers to leverage the power of GPUs without the need for expensive hardware. Key features: - Easy GPU function definition - Run functions in the cloud with a single decorator - Supports GPU scaling and dependency management - Integrates with the Runpod API and serverless worker SDK Tech stack: Python Why it matters: This library simplifies the process of using GPUs in the cloud, making it accessible to developers without requiring significant infrastructure setup. It provides a straightforward way to run GPU-accelerated functions and integrates seamlessly with the Runpod ecosystem, enabling developers to deploy and scale their applications efficiently.

### Overview: ShadowKV: KV Cache in Shadows for High-Throughput Long-Context LLM Inference ShadowKV is a novel KV cache system designed specifically for high-throughput long-context language model (LLM) inference. Leveraging the power of shadows, this project aims to accelerate inference by efficiently caching frequently accessed data. The system is optimized for large-scale models, ensuring seamless integration with existing frameworks and minimizing overhead. ### Key Features: - **High Throughput**: Utilizes shadows to achieve high throughput by reducing the number of cache misses. - **Long-Context Support**: Designed for models with long context lengths, ensuring optimal performance for complex queries. - **Training-Free**: Does not require any additional training, making it a lightweight solution. - **Scalability**: Easily scalable to accommodate different hardware configurations and model sizes. ### Tech Stack: Python ShadowKV is implemented using Python, leveraging the power of the Python ecosystem for efficient data handling and model inference. ### Why It Matters: ShadowKV offers a significant improvement in inference performance, particularly for models with long context lengths. By reducing the number of cache misses and optimizing memory access patterns, it enables faster and more efficient inference, making it a valuable tool for high-throughput LLM inference. This project is particularly relevant for applications requiring high performance and low latency, such as real-time language processing and large-scale language models. ### Repository: ByteDance-Seed/ShadowKV The project is hosted on GitHub, with a total of 316 GitHub stars, indicating its growing interest and importance in the field of

### Overview: ComfyUI-Qwen3-TTS is a custom node suite designed for the ComfyUI platform, tailored for the Qwen3-TTS model. It supports a wide range of features including custom voices, voice design, voice cloning, and fine-tuning, making it ideal for voice synthesis and customization. ### Key Features: - **ComfyUI Model Folder Integration**: Models are stored in `ComfyUI/models/Qwen3-TTS/`, keeping your models organized alongside other ComfyUI models. - **On-Demand Download**: Only downloads the model you select—no need to pre-download all variants. - **Custom Voice**: Use 9 preset high-quality voices (Vivian, Ryan, etc.). - **Voice Design**: Create new voices using natural language descriptions. - **Voice Cloning**: Clone voices from a short reference audio clip. - **Fine-Tuning**: Train a custom voice model using your own dataset (folder of .wav + .txt files). - Resume training from checkpoints - VRAM optimizations: gradient checkpointing, 8-bit AdamW, configurable batch sizes - Per-epoch checkpointing with automatic cleanup - Support for both 1.7B and 0.6B models - **Audio Comparison**: Evaluate fine-tuned models with speaker similarity and mel spectrogram metrics. - **Cross-Lingual Support**: Generate speech in Chinese, English, Japanese, Korean, German, French, Russian, Portuguese, Spanish, and Italian. - **Flexible Attention**: robust support for `flash_attention