Projects

### Overview: This repository is a curated list of research papers and resources on computational linguistic approaches to code-switching research. The list will be updated over time and welcomes contributions from the community. ### Key Features: - **Workshops**: Organized at major conferences like NAACL, EMNLP, and ACL. - **Research Papers**: Includes surveys, large language models, language identification and POS tagging, corpus studies, language modeling, speech recognition, discourse analysis, generation, speech synthesis, metrics, representation learning, machine translation, speech translation, natural language understanding, named entity recognition, linguistics, affective computing, dialog and conversational systems, discourse, syntax, adversarial attack, social linguistics, benchmarking, social media, text normalization, and toolkits. - **Books and Theses**: Contributions from researchers and students in the field of code-switching. ### Tech Stack: Unknown. ### Why It Matters: This repository serves as a comprehensive resource for researchers and practitioners interested in computational linguistic approaches to code-switching. It provides a platform to share and discuss the latest advancements and challenges in the field, fostering collaboration and knowledge exchange.

Overview: This repository contains the official implementation of 3D Equivariant Diffusion for Target-Aware Molecule Generation and Affinity Prediction, presented at ICLR 2023. Key features: - 3D Equivariant Diffusion model for generating molecules and predicting affinities - Data preprocessing and training from scratch using lmdb and PyTorch Geometric - CrossDocked2020 dataset for training and testing - Vina Docking integration for molecular docking Tech stack: Python, PyTorch, PyTorch Geometric, RDKit, OpenBabel, TensorBoard, EasyDict, LMDB Why it matters: This model and dataset represent significant advancements in target-aware molecule generation and affinity prediction, offering a powerful tool for chemists and researchers in the field of drug discovery and materials science.

### Overview: Neural Speed is an innovative library designed to efficiently infer large language models (LLMs) on Intel platforms. It leverages the latest low-bit quantization techniques from Intel Neural Compressor, offering up to 40x performance improvements over Llama.cpp. The library is optimized for CPUs with support for ISAs like AMX, VNNI, AVX512F, AVX_VNNI, and AVX2, making it highly efficient for inference tasks. ### Key Features: - **Highly Optimized Kernels**: The library features optimized kernels for N-bit weight quantization, including support for int1, int2, int3, int4, int5, int6, int7, and int8. - **Performance Speedup**: Achieves up to 40x performance improvement on popular LLMs compared to Llama.cpp. - **Tensor Parallelism**: Supports tensor parallelism across sockets/nodes on CPUs for efficient inference. ### Tech Stack: - **Intel Neural Compressor**: Powered by the latest low-bit quantization techniques. - **Intel Xeon Scalable Processors**: Supported hardware. - **Intel Xeon CPU Max Series**: Supported hardware. - **Intel Core Processors**: Supported hardware. ### Why It Matters: Neural Speed is designed to optimize the inference of large language models on Intel platforms, making it an ideal choice for applications requiring high performance and low power consumption. The library's performance improvements and efficient parallel processing capabilities make it a valuable tool for developers working on AI and machine learning projects on Intel hardware

Overview: Megabots is a state-of-the-art, production-ready language model library designed to simplify the creation of large language models, making it easy for developers to build bots without the need for extensive coding. Key features: - 🚀 Production-ready, easy-to-use LLM apps - 📂 Answers questions over documents - 📱 Connects to vector databases - 🏓 Automatically expose the bot as a robust API using FastAPI - 🎖️ Automatically expose the bot as a UI using Gradio Tech stack: Python Why it matters: Megabots simplifies the process of creating and deploying large language models, making it accessible and efficient for developers without requiring extensive programming knowledge.

Overview: 医药知识图谱自动问答系统实现项目，包括构建知识图谱、基于知识图谱的问答系统以及前端交互。 Key features: - 实体识别（基于词典+BERT_CRF） - 实体链接（Sentence-BERT匹配） - 意图识别（基于提问词+领域词词典） Tech stack: JavaScript Why it matters: 该项目通过爬虫技术构建了详细的医药知识图谱，并实现了基于知识图谱的问答系统，为医药领域的信息检索和知识应用提供了强大的工具。

### Overview: ThinkRAG is a LLM RAG system that can be easily deployed on a laptop to implement Q&A with local knowledge base. This system is built on LlamaIndex and Streamlit, and has been optimized for Chinese users in various fields such as model selection and text processing. ### Key Features: - Complete application of the LlamaIndex framework - Development mode supports local file storage without the need to install any databases - No GPU support is required to run on a laptop - Supports locally deployed models and offline use Specifically, ThinkRAG has also made a lot of customizations and optimizations for Chinese users: - Uses Spacy text splitter for better handling of Chinese characters - Employs Chinese title enhancement features - Uses Chinese prompt templates for Q&A and refinement processes - Default support for China LLM service providers such as DeepSeek, Moonshot, and ZhiPu - Uses bilingual embedding models, such as bge-large-zh-v1.5 from BAAI ### Tech Stack: Python ### Why it Matters: ThinkRAG is a LLM application developed for professionals, researchers, students, and other knowledge workers, which can be used directly on a laptop with all knowledge and data stored locally on the computer. It supports the deployment of LLMs locally, making it easy to use and deploy, and it is committed to creating an application system that is directly usable, useful, and easy to use.

Overview: This is a TensorFlow implementation of the Attention Is All You Need (2017) model, focusing on the Transformer architecture. It includes a detailed project structure, requirements, and usage instructions for a developer audience. Key features: - Transformer architecture implementation - Support for different datasets and configurations - Training and evaluation capabilities Tech stack: Python Why it matters: This project showcases a practical application of the Transformer model, demonstrating its effectiveness in sequence-to-sequence tasks, particularly in language translation.

Overview: This Node.js interface allows developers to utilize Google's word2vec tool for efficient NLP tasks. Key features: - Node.js interface - Efficient implementation of continuous bag-of-words and skip-gram architectures - Can be used for various NLP applications Tech stack: C Why it matters: This tool provides a powerful way to represent and manipulate words in a vector space, which is crucial for many NLP tasks such as text classification, clustering, and more. Repository: Planeshifter/node-word2vec Primary language: C GitHub stars: 349

### Overview: MarkEverythingDown is a versatile document conversion tool designed to streamline the process of converting various file formats into clean, structured markdown. It leverages powerful multimodal LLMs, including Qwen2.5 VL, to ensure high-quality processing of visual content while being user-friendly and accessible. ### Key Features: - **Multi-format support**: Convert PDFs, DOCX, PPTX, XLSX, images, code files, notebooks, and markdown variants. - **Intelligent processing**: Automatically selects the appropriate processor for each file type. - **Vision AI support**: Optimized for Qwen2.5 VL models with OpenAI-compatible interface. - **Dual processing options**: Support local inference APIs and cloud APIs. - **Batch processing**: Process multiple files at once with a simple interface. - **User-friendly UI**: Easy-to-use web UI with Gradio and helpful tooltips. - **Command line interface**: Quick conversions from the terminal. ### Tech Stack: MarkEverythingDown is built using Python, leveraging Qwen2.5 VL models through OpenAI-compatible APIs for processing both local and cloud-based options. ### Why It Matters: This tool simplifies the conversion of documents into markdown format, making it LLM-friendly and easy for humans to read. It is designed to streamline the conversion of documents into markdown format, ensuring seamless integration with RAG applications and SFT dataset preparations.

### Overview: Kaldi Active Grammar is a Python package designed for context-based command & control of computer applications using the Kaldi automatic speech recognition engine. It enables the dynamic activation and inactivity of grammars at decode-time, improving accuracy by selectively recognizing only the relevant grammars. ### Key Features: - **Dynamic Grammar Activation:** Each grammar can be marked as active or inactive per-utterance, allowing for selective recognition. - **Shared Dictation Grammar:** The dictation grammar can be shared across multiple command grammars, compiled quickly without including large-vocabulary dictation directly. - **Multiple Independent Grammars:** Multiple grammars can be compiled separately and stitched together dynamically at decode-time. - **Monolithic Kaldi Decoding:** Kaldi decoding graphs are monolithic and require expensive upfront compilation, but Kaldi's new grammar framework allows dynamic compilation and stitching. ### Tech Stack: Python ### Why It Matters: This project extends the capabilities of Kaldi by enabling dynamic, per-utterance activation of grammars, leading to improved accuracy and reduced computational overhead. The shared dictation grammar further enhances flexibility and efficiency in application control.

### Overview: This repository contains a comprehensive collection of academic publications and methodologies related to the classification of Code Large Language Models' pre-training tasks, downstream tasks, and their application in the field of Software Engineering (LLM4SE). The survey covers a wide range of topics, including collaborative human-AI approaches, code-mixed language models, fuzzing robotic software, and more. ### Key Features: - **Collaborative Human-AI Approach to Python Package Creation** (PyGen) - **Code-mixed LLM: Improve Large Language Models' Capability to Handle Code-Mixing through Reinforcement Learning from AI Feedback** - **Fuzzing Robotic Software Using High-Performance Computing (HPC)** - **InfiBench: Evaluating the Question-Answering Capabilities of Code Large Language Models** - **Transforming the Field of Vulnerability Prediction: Are Large Language Models the Key?** - **IaC-Eval: A Code Generation Benchmark for Cloud Infrastructure-as-Code Programs** - **A Comprehensive Survey of AI-Driven Advancements and Techniques in Automated Program Repair and Code Generation** - **Towards the Integration of Large Language Models and Automatic Assessment Tools: Enhancing Student Support in Programming Assignments** - **Semantic Error Detection in Code Translation Using Knowledge-Driven Static Analysis with AI Chain** - **RUG: Turbo LLM for Rust Unit Test Generation** - **ROCODE: Integrating Backtracking Mechanism and Program Analysis in Large Language Models for Code Generation** - **Automatically Write Code Checker: An LLM-based Approach with Logic-guided

### Overview: This repository contains an impressive collection of over 500 AI market maps from 2026, 2025, and 2024, curated by Joy Larkin. The repository serves as a valuable resource for AI enthusiasts and industry professionals, providing a comprehensive overview of the latest trends and developments in the AI market. ### Key Features: - **Market Maps**: Detailed visual representations of AI market landscapes. - **Data Sources**: Curation from top VCs, industry analysts, and AI practitioners. - **Geographical Focus**: Maps are categorized by region, ensuring a global perspective. - **Up-to-Date**: Regular updates ensure the repository remains relevant and accurate. ### Tech Stack: The repository is built using Python, making it accessible and easy to maintain. ### Why it Matters: This repository is crucial for anyone interested in the AI market, providing a wealth of information on the latest trends, market maps, and AI industry positioning. It serves as a valuable resource for both beginners and experienced professionals in the field. ### Repository: The repository is hosted on GitHub and has received 347 GitHub stars, indicating its popularity and usefulness among the developer community. ### Methodology: The data collection workflow includes using Claude, Exa Websets, and human-in-the-loop curation for "taste," with some additional methods not listed. The data is organized by quarter to track the evolving trends in the AI market. ### Notes: - The README.md file is regularly updated with new market maps, and the CSV file

Overview: This implementation of Soft MoE, proposed by Brain's Vision team, is designed for non-autoregressive encoders and explores its potential for text-to-image models. It aims to address the limitation of fixed number of slots and dynamic slot management. Key features: - Implements Soft MoE for Pytorch - Supports non-autoregressive encoders - Exploring potential for autoregressive encoders - Dynamic slot management to address limitations - Provides a way to add a transformer in place of a feedforward at a certain layer Tech stack: Python Why it matters: This implementation offers a flexible solution for MoE in Pytorch, potentially improving performance in text-to-image models and addressing the limitations of fixed slot management. Repository: lucidrains/soft-moe-pytorch GitHub stars: 349

Overview: 이 저장소는 텍스트 요약 분야의 주요 연구 주제, 최신 논문, 사용 가능한 모델 및 데이터를 추천 자료와 함께 요약 모델을 정리한 공간입니다. Key features: - 텍스트 요약의 정의와 주요 구분 - 최신 논문 추천 - 요약 모델 및 데이터셋 목록 - 다양한 요약 방법과 주제 분류 Tech stack: Python, PyTorch, TensorFlow 등 Why it matters: 이 저장소는 텍스트 요약 분야의 주요 연구 주제, 최신 논문, 사용 가능한 모델 및 데이터를 정리하여 요약 모델을 이해하고 활용하는 데 도움이 될 수 있습니다.

Overview: This repository contains code for PIXEL, a language model that operates on text rendered as images, removing the need for a fixed vocabulary. It is pretrained on English Wikipedia and BookCorpus, outperforming BERT on non-pretrained scripts. Key features: - Pretrained on English Wikipedia and BookCorpus - Monolingual model for language processing - Pixel-based encoder and decoder architecture Tech stack: Python, PyTorch, Transformers, ViT-MAE Why it matters: PIXEL demonstrates the potential of image-based language models, potentially enabling more versatile and context-aware language processing. Repository: xplip/pixel GitHub stars: 348 GitHub description: Research code for pixel-based encoder of language (PIXEL) README excerpt: This repository contains code for **PIXEL**, the **Pix**el-based **E**ncoder of **L**anguage. PIXEL is a language model that operates on text rendered as images, fully removing the need for a fixed vocabulary. This effectively allows for transfer to any language and script that can be typeset on your computer screen. We pretrained a monolingual PIXEL model on the English Wikipedia and BookCorpus (in total around 3.2B words), the same data as BERT, and showed that PIXEL substantially outperforms BERT on syntactic and semantic processing tasks on scripts that are not found in the pretraining data, but PIXEL is slightly weaker than BERT when working with Latin scripts. For details about PIXEL, please have a look at our paper [Language Modelling with Pixels](https

### Overview: gflow-cli is a Python CLI and MCP server designed to drive Google Flow from the command line. It allows you to script, batch, and pipeline Veo video and Imagen image generation. The CLI is designed to be fully programmatically controlled, giving you and your agents full programmatic control over the generation process. ### Key Features: - **Batch Generation:** Loop prompts straight from the shell: `for p in $(cat prompts.txt); do gflow image t2i "$p"; done`. - **Image and Video Generation:** Drive Veo video and Imagen image generation. - **Pipeline-Ready:** Generates images and videos in a pipeline-ready format. ### Tech Stack: Python ### Why It Matters: gflow-cli provides a powerful and flexible way to automate Google Flow, allowing you to batch and script your image and video generation tasks. It gives you full control over the generation process, making it ideal for batch workflows and pipeline-ready projects.

Overview: This repository is a curated list of awesome Multimodal studies, focusing on foundational models, visual understanding, omni understanding, unified understanding and generation, coding/gui, diffusion mllm, multimodal embedding/retrieval, image understanding benchmark, video understanding benchmark, audio, multimodal dialogue, multimodal learning, image generation, video generation, multimodal dataset, and multimodal survey. Key features: - Foundation Model (Textual and Multimodal) - Visual Understanding - Omni Understanding - Unified Understanding and Generation - Coding / GUI - Diffusion MLLM - Multimodal Embedding/Retrieval - Image Understanding Benchmark - Video Understanding Benchmark - Audio - Multimodal Dialogue - Multimodal Learning - Image Generation - Video Generation - Multimodal Dataset - Multimodal Survey Tech stack: Python, Hugging Face, and other relevant libraries and frameworks. Why it matters: This repository serves as a valuable resource for researchers and developers interested in Multimodal studies, providing curated papers and resources that can accelerate the development of cutting-edge technologies in the field.

### Overview: Program-Y is a Python-based AIML 2.0 chatbot framework designed to simplify the creation of AI-powered chatbots. ### Key Features: - **Compliance with AIML 2.1**: Ensures seamless integration with the AIML standard. - **Configuration Engine**: Simplifies OOB and Trigger configuration. - **Unit Tests**: Over 2000 unit tests for 99% coverage, improving code quality. - **Configuration Structure Refactoring**: Aligns with bot architecture, making config files easier to manage. - **Embeddable Clients**: Simplifies adding a bot to Python apps with just 2 lines of code. - **Unified Naming Convention**: Renames configuration items to use underscores for better readability. ### Tech Stack: Python 3.x ### Why it Matters: Program-Y provides a comprehensive solution for developers looking to create AI-driven chatbots, making the process more manageable and efficient. The framework includes robust features for configuration, unit testing, and embeddable clients, ensuring a high-quality and sustainable development experience.

Overview: Betty is a library designed for generalized meta-learning and multilevel optimization, enabling simple and modular implementation of large-scale applications such as meta-learning, hyperparameter optimization, neural architecture search, and data reweighting. Key features: - Distributed Data Parallel (DDP) - ZeRO Redundancy Optimizer (ZeRO) - Fully Sharded Data Parallel (FSDP) - Multilevel optimization - Hierarchical problem structure - Simple and modular programming interface Tech stack: Python Why it matters: Betty simplifies the process of implementing large-scale meta-learning and optimization problems by providing a simple and modular programming interface. It allows users to easily scale up meta-learning and even meta-meta-learning by simply changing one line of code. Repository: leopard-ai/betty Primary language: Python GitHub stars: 348 GitHub description: Betty: an automatic differentiation library for generalized meta-learning and multilevel optimization README excerpt: An automatic differentiation library for generalized meta-learning and multilevel optimization<br> <a href="https://leopard-ai.github.io/betty/">Docs</a> | <a href="https://leopard-ai.github.io/betty/tutorial/basic/basic.html">Tutorials</a> | <a href="https://github.com/leopard-ai/betty/tree/main/examples">Examples</a> | <a href="https://openreview.net/pdf?id=LV_MeMS38Q9">Paper</a> | <a href="https://github.com/le

### Overview: TensorLayer is a powerful library for implementing deep learning models, especially for researchers and developers. It simplifies the process of building, training, and deploying neural networks, making it easier to experiment with various architectures and techniques. ### Key Features: - **Model Definition**: TensorLayer provides a simple syntax for defining models, making it easier to build complex architectures. - **Model Compilation**: It allows for easy compilation of models into TensorFlow graphs, facilitating efficient computation. - **Model Optimization**: TensorLayer includes various optimization techniques, such as dropout, batch normalization, and more, to improve model performance. - **Model Training**: It supports both training and testing phases, with the ability to switch between them using the `all_drop` layer. - **Model Sharing**: Models can be shared between training and testing phases, allowing for different configurations for each. ### Tech Stack: - **TensorFlow**: The underlying framework used for building and training models. ### Why it Matters: TensorLayer simplifies the process of implementing and training deep learning models, making it an essential tool for researchers and developers in the field of machine learning. Its features and techniques make it easier to experiment with different architectures and optimize models for better performance.