Projects

### Overview: LTP 是一个全面的中文自然语言处理平台，涵盖了分词、词性标注、句法分析等核心功能。 ### Key Features: - **分词与词性标注**: 实现了精确的中文分词和词性标注，支持多种语言模型。 - **句法分析**: 提供了基于规则和深度学习的句法分析工具。 - **多任务学习**: 采用多任务学习框架，共享预训练模型，提高跨任务性能。 ### Tech Stack: Python ### Why it Matters: LTP 是一个开源的中文自然语言处理平台，适用于各种中文文本处理任务，如分词、词性标注、句法分析等，为开发者提供了高效便捷的工具。 Repository: HIT-SCIR/ltp Primary language: Python GitHub stars: 5264 GitHub description: Language Technology Platform README excerpt:   

Overview: WeNet is a production-first and production-ready end-to-end speech recognition toolkit designed for accurate and lightweight solutions. Key features: - Accurate results on various public speech datasets - Easy installation and usage, well-documented - Production-first design principles Tech stack: Python Why it matters: WeNet offers robust, production-ready speech recognition capabilities with high accuracy and efficiency, making it suitable for real-world applications. Repository: wenet-e2e/wenet GitHub stars: 5247

### Overview: XTuner V1 is a next-generation LLM training engine designed specifically for ultra-large-scale MoE models. It offers scalable training without the need for complex parallelism, optimized memory efficiency, and superior performance across various scenarios. ### Key Features: - **Dropless Training:** Scalable without the need for expert parallelism, enabling training of 200B-scale MoE models without the need for expert parallelism. - **Long Sequence Support:** Train 200B MoE models on 64k sequence lengths without sequence parallelism, supported by advanced memory optimization techniques. - **Flexible Scaling:** Full support for DeepSpeed Ulysses sequence parallelism, with linearly scalable maximum sequence length. - **Robust Performance:** Maintains stability during long sequence training, even with expert load imbalance. ### Tech Stack: Python ### Why it Matters: XTuner V1 is crucial for training ultra-large-scale MoE models efficiently, offering a significant improvement in training efficiency and performance compared to traditional 3D parallel architectures. Repository: InternLM/xtuner

### Overview: Kiln is a comprehensive AI development platform that includes evaluations, automatic optimization, prompts, RAG, fine-tuning, synthetic data generation, agents, and more, all accessible through a free desktop app and open-source Python library. ### Key Features: - **Intuitive App:** Easy-to-use desktop app for Mac, Windows, and Linux, with one-click install. - **Eval Builder:** Auto-generate evaluations (judge + synthetic eval dataset) in 10 minutes. - **Auto-Optimize:** Automatically find the best way to run AI tasks, optimizing prompt, model, tools, skills, subagents, parameters, and more. - **AI Assistant:** Your AI data-science partner, proposing improvements, optimizing prompts, running experiments, creating evals, and more. - **Git-Native Collaboration:** Syncs with Git, even for teammates without Git knowledge. - **RAG (Relevance-Augmented Generation):** Drag-and-drop docs to create a RAG, auto-generated RAG evals from your own documents. - **Subagents:** Create and manage subagents for complex AI tasks. - **Fine-Tuning:** Fine-tune models and agents for improved performance. - **Synthetic Data Generation:** Generate synthetic data for training and testing AI models. ### Tech Stack: Python ### Why it Matters: Kiln simplifies the entire AI development process, from evaluation and optimization to collaboration and agent creation, making it an essential tool for anyone looking to build and ship better AI systems. ### GitHub Stars: 5,197 ###

Overview: DigUp is a powerful tool that allows you to search your computer by the contents of your files, including photos, PDFs, documents, recordings, videos, and code. It uses Google DeepMind's EmbeddingGemma 2 to understand and match the content of your files. Key features: - Searches for files by their content, not just by their name. - Supports multiple languages, including English, Bengali, Arabic, and many more. - Can find exact words and phrases within files. - Tracks and indexes your files, ensuring privacy and security. Tech stack: Swift Why it matters: DigUp is a valuable tool for anyone who needs to quickly find specific information on their computer, whether it's for work, personal projects, or everyday tasks. Its ability to search by content makes it incredibly useful for managing files and finding relevant information quickly and efficiently. Repository: ARahim3/DigUp Primary language: Swift GitHub stars: 48,140 GitHub description: Search your computer by what's in your files: photos, PDFs, documents, recordings, videos and code. On-device on Mac, Windows and Linux, built on Google DeepMind's EmbeddingGemma 2.

Overview: EasyR1 is an efficient, scalable, multi-modality RL training framework based on veRL, designed to support vision language models and various vision-text datasets. It includes models like Llama3/Qwen2/Qwen2.5/Qwen3 and vision-language models like Qwen2-VL/Qwen2.5-VL/Qwen3-VL, along with algorithms like GRPO, DAPO, Reinforce++, ReMax, RLOO, GSPO, and support for custom datasets and tricks such as padding-free training, LoRA training, and resuming from checkpoints. Key features: - Supported models: Llama3/Qwen2/Qwen2.5/Qwen3, Qwen2-VL/Qwen2.5-VL/Qwen3-VL, DeepSeek-R1 distill models - Supported algorithms: GRPO, DAPO, Reinforce++, ReMax, RLOO, GSPO - Supported datasets: Any text, vision-text dataset in a specific format - Supported tricks: Padding-free training, LoRA training, resuming from the latest/best checkpoint, Wandb & SwanLab & Mlflow & Tensorboard tracking Tech stack: Python Why it matters: EasyR1 is designed to efficiently and scalably train vision language models, making it a valuable tool for researchers and developers working on large language models and vision-language tasks. Repository: hiyouga/EasyR1 Primary language: Python GitHub stars: 5198 GitHub description: EasyR1: An Efficient, Scal