Projects

### Overview: This project provides the latest versions of Disco Diffusion, including Turbo and 3D animation, running locally on Windows instead of Google Colab. It includes detailed instructions on setting up the environment, including requirements for a Nvidia GPU with at least 8GB VRAM and Windows 10/11. The project also includes an experimental batch mode for creating multiple videos with different prompts. ### Key Features: - Latest Disco Diffusion versions with Turbo and 3D animation. - Instructions for running Disco Diffusion locally on Windows. - Batch mode for creating multiple videos with different prompts. - Detailed setup guide for Windows environment. ### Tech Stack: - Disco Diffusion: Latest versions with Turbo and 3D animation. - Anaconda: Used for managing packages and creating conda environments. - WSL2: Windows Subsystem for Linux (Ubuntu environment). - Nvidia GPU: For running Disco Diffusion with Turbo and 3D animation. ### Why It Matters: This project is essential for developers and artists who need to work with Disco Diffusion locally, ensuring compatibility and performance without relying on Google Colab. The batch mode allows for quick creation of multiple videos with different prompts, streamlining the workflow and saving time.

### Overview: Anti-DreamBooth is a security system designed to protect users from personalized text-to-image synthesis, specifically DreamBooth. It introduces subtle noise perturbation to user images before they are published, making it difficult for DreamBooth models trained on perturbed images to generate realistic images of any individual. ### Key features: - **Subtle Noise Perturbation**: Adds noise to user images to disrupt the quality of DreamBooth-generated images. - **Wide Range of Algorithms**: Investigates and evaluates various perturbation optimization algorithms. - **Facial Datasets**: Tested on two facial datasets for robust performance. - **Versatile Models**: Extensively evaluated on different versions of text-to-image models. ### Tech stack: Python ### Why it matters: Anti-DreamBooth is crucial in safeguarding users from the malicious use of DreamBooth, which can produce fake news or disturbing content targeting any individual victim. It ensures privacy and security by preventing such misuse, thereby mitigating the severe negative social impact of such technology.

### Overview: This repository contains a Persian swear word dataset that can be used in various programming languages to filter unwanted content. The dataset is available in JSON format and can be easily integrated into production applications. ### Key Features: - **Java**: Provides a class `PersianSwear` with methods to add, remove, check, and filter bad words in text. - **Go**: Offers a function `PersianSwear.go` for the same functionalities. - **PHP**: Provides a class `PersianSwear.php` with similar methods for adding, removing, checking, and filtering bad words in text. - **Python**: Not explicitly mentioned in the README, but Python support is available. - **JavaScript**: Not explicitly mentioned in the README, but JavaScript support is available. - **TypeScript**: Not explicitly mentioned in the README, but TypeScript support is available. - **C#**: Provides a class `PersianSwear` with methods to add, remove, check, and filter bad words in text. - **Swift**: Not explicitly mentioned in the README, but Swift support is available. ### Tech Stack: The repository uses C# as the primary programming language. ### Why It Matters: This dataset can be used in various programming languages to filter out unwanted content, ensuring that the text is clean and appropriate for the intended audience. The dataset is available in JSON format, making it easy to integrate into existing projects or new ones. ### GitHub Stars: The repository has received 316 GitHub stars, indicating significant interest and use by developers. ### Installation: -

### Overview: Shodh-Memory is a local, offline Rust binary designed for AI agents and robots. It learns from usage, forgets irrelevant data, and strengthens what matters. No LLM calls, no external services, and no cloud storage. Works with chat agents, robots, and edge devices. ### Key Features: - **Persistent Memory**: Stores and remembers what matters. - **Self-Improving Learning**: Learns from usage and adapts over time. - **Decay and Forget**: Automatically forgets irrelevant context. - **Deterministic and Auditable**: No cloud, no API keys, no external databases. ### Tech Stack: Rust ### Why It Matters: - **No LLM Calls**: Eliminates the need for expensive LLM API calls. - **Deterministic and Auditable**: Ensures consistent and traceable behavior. - **Offline Operation**: Works fully offline, eliminating the need for cloud storage or API keys. - **Robotic Compatibility**: Native support for robotics and ROS2. - **Binary Size**: Compact, just over 17MB. ### Why Not Just Use mem0 / Cognee / Zep? - **LLM Calls**: 0 vs 2+ per add, 3+ per cognify, 2+ per episode. - **External Services**: None vs OpenAI + vector DB, Neo4j + vector DB, OpenAI + Neo4j. - **Time to Store**: 55ms vs ~20 seconds, seconds, seconds. - **Learns from Usage**: Yes

### Overview: This repository provides a PyTorch implementation of the Learning to Cluster (L2C) strategy, which is a deep clustering approach based on transfer learning. It includes two learning criteria: Meta Classification Likelihood (MCL) and KLD-based Contrastive Loss (KCL). The repository supports PyTorch 1.0, Python 2.7, 3.6, and 3.7, and is based on the work presented in the blog post "Learning to Cluster" and the conference papers "Learning to Cluster in order to Transfer Across Domains and Tasks" and "Neural network-based clustering using pairwise constraints". ### Key features: - PyTorch implementation of L2C strategy - Two learning criteria: MCL and KCL - PyTorch 1.0, Python 2.7, 3.6, and 3.7 support - Includes examples for supervised classification/clustering and unsupervised clustering (cross-task transfer learning) ### Tech stack: PyTorch ### Why it matters: This repository provides a practical tool for researchers and practitioners interested in deep clustering and transfer learning. It offers a way to apply deep learning techniques to clustering tasks, which is not commonly used in the field. The implementation is based on the original work and includes utilities that are different from the Lua-based implementation used in the reference papers, ensuring a fair comparison with the original results.

Overview: This project is an official PyTorch implementation of "BERT-of-Theseus: Compressing BERT by Progressive Module Replacing" paper presented at EMNLP 2020. The code aims to compress BERT by progressively replacing its components, making it more efficient for various NLP tasks. Key features: - Progressive module replacing - Compresses BERT for better performance - Supports GLUE tasks and leaderboard submission Tech stack: PyTorch Why it matters: This project is significant for researchers and developers looking to optimize BERT for specific NLP tasks, potentially improving model efficiency and performance on downstream tasks like GLUE. Repository: JetRunner/BERT-of-Theseus Primary language: Python GitHub stars: 315 GitHub description: The official PyTorch implementation for "BERT-of-Theseus: Compressing BERT by Progressive Module Replacing" (EMNLP 2020). README excerpt: This repository contains the code for the paper "BERT-of-Theseus: Compressing BERT by Progressive Module Replacing". It provides instructions on how to fine-tune a predecessor model, compress BERT using a replacement scheduler, and run predictions on GLUE tasks. Additionally, it includes a script for preparing predictions for leaderboard submission. The code is built on the Hugging Face Transformers library and supports a 6-layer pretrained model on MNLI, outperforming DistillBERT on six tasks of GLUE.

### Overview: Vocalis is a sophisticated AI assistant designed for seamless, low-latency speech-to-speech communication. Built on a modern React frontend and a FastAPI backend, it offers a responsive, low-latency conversational experience with advanced visual feedback. ### Key Features: - **Barge-In Interruption**: Interrupt the AI mid-speech for a truly natural conversation experience. - **AI-Initiated Greetings**: Assistant automatically welcomes users with a contextual greeting. - **Intelligent Follow-Ups**: System detects silence and continues conversation with natural follow-up questions. - **Conversation Memory**: Maintains context throughout the conversation session. - **Contextual Understanding**: Processes conversation history for coherent, relevant responses. - **Image Analysis**: Upload and discuss images with integrated visual understanding. - **Session Management**: Save, load, and manage conversation sessions with customisable titles. ### Tech Stack: TypeScript ### Why it Matters: Vocalis provides a sophisticated AI assistant with speech-to-speech capabilities that are built on a modern React frontend with a FastAPI backend. It offers a responsive, low-latency conversational experience with advanced visual feedback, making it ideal for applications requiring real-time interaction and seamless integration with local LLM/TTS services.

### Overview Prompthon Agentic Labs publishes the Agent Systems Handbook by Prompthon: an AI-native field guide for students, practitioners, and builders exploring modern agent systems from different angles. Built on **learn, question, and innovate**, the lab is shaped by learners and grounded in real industry practice. It helps readers understand the space, apply AI effectively, or build real systems through parallel paths rather than a single track. ### Key Features - **LangGraph**: A language for describing and reasoning about AI systems. - **MCP/A2A Interoperability**: Standardized communication protocols for integrating AI systems. - **Context Engineering**: Techniques for managing and understanding the context in which AI systems operate. - **Agent Memory**: Persistent storage and retrieval of agent state and interactions. - **Evaluation**: Metrics and methods for assessing the performance of AI agents. - **Observability**: Tools and techniques for monitoring and logging agent interactions. - **Multi-Agent Architecture**: Design patterns and best practices for building complex AI systems. ### Tech Stack - **MDX**: Markdown Documentation ### Why It Matters This handbook provides a comprehensive overview of modern agent systems, covering essential concepts and tools for AI-native learners, practitioners, and builders. It helps readers understand the space, apply AI effectively, or build real systems through parallel paths rather than a single track. The lab is designed for different kinds of learners and different intentions, from broad understanding and trend awareness to applying AI tools to daily work and building real systems.

Overview: Chatette is a powerful dataset generator for Rasa NLU, inspired by Chatito, designed to easily generate datasets for Natural Language Understanding tasks without much effort. Key features: - Generates datasets for Rasa NLU - Uses a Domain Specific Language (DSL) for easy template definition - Supports interactive mode for quick experimentation Tech stack: Python Why it matters: Chatette simplifies the process of creating large datasets for NLU tasks, making it easier for developers to test and refine their models without needing extensive manual data entry. Repository: SimGus/Chatette Primary language: Python GitHub stars: 315 GitHub description: A powerful dataset generator for Rasa NLU, inspired by Chatito README excerpt: A data generator for Rasa NLU, designed to easily generate datasets for Natural Language Understanding tasks without too much of a headache.

### Overview: 本项目是一个中文智能客服机器人demo，包含闲聊和专业问答2个部分，支持自定义组件。系统采用基于tf-idf检索的召回模型和基于CNN的精排模型，旨在构建一个精确且快速的客服聊天机器人。 ### Key features: - **基于tf-idf的召回模型** - **基于CNN的精排模型** - **闲聊模式和专业问答模式** - **自定义组件支持** ### Tech stack: Python ### Why it matters: 该系统能够帮助开发者快速构建一个功能强大且易于维护的智能客服机器人，适用于多种应用场景，尤其在需要快速响应和准确信息提取的环境中表现突出。

Overview: This project offers a comprehensive guide to understanding how Llama and its components operate in practice, focusing on practical implementation details rather than theoretical facts. It includes Go code and detailed documentation. Key features: - Go code implementation - Detailed documentation - CLI application for executing predefined or custom prompts Tech stack: Go Why it matters: This project provides a practical understanding of Llama and its components without relying on external libraries, making it ideal for educational and experimental purposes. Repository: adalkiran/llama-nuts-and-bolts GitHub stars: 315

Overview: AI Engineering Lab is a comprehensive 24-week AI engineering course designed to help beginners transition from Python to production-grade AI systems. This program covers topics such as machine learning, large language models, prompt and context engineering, and more, with a focus on hands-on learning and practical application. Key features: - Continuous case study: Freight operations case study across 24 weeks. - Runnable notebooks: Python, SQL, and PySpark for hands-on practice. - Metrics and error notes: From Week 3 onwards, every artifact must include a metric and error note. - Curriculum: Developed by Zorost Intelligence AI Lab, MIT licensed, and available for free. Tech stack: Python, Jupyter Notebook, SQL, PySpark, Git, GitHub, and various cloud services. Why it matters: This program provides a structured, hands-on learning experience for software developers, analysts, and students interested in AI engineering, ensuring practical application and real-world problem-solving skills. Repository: zorost/AI-Engineering-Lab Primary language: Jupyter Notebook GitHub stars: 319 GitHub description: A free, self-paced 24-week AI engineering course: Python, machine learning, LLMs, RAG, fine-tuning, agents and MCP, Azure and Vertex and Bedrock, and Databricks. 43 runnable notebooks, one continuous case study. MIT licensed, no signup. By Zorost Intelligence AI Lab.

### Overview: 这是一个存放北京理工大学计算机科学与技术专业相关课程资料的仓库，适合计算机科学与技术专业的学弟学妹参考。 ### Key features: - 课程大纲和历年试卷 - 课程属性明确标识 - 课程名称和考核方式详细列出 ### Tech stack: VHDL ### Why it matters: 这个仓库为计算机科学与技术专业的学生提供了丰富的课程资料和学习经验，帮助他们在学习过程中更好地理解和掌握专业知识。

Overview: RAGLAB is a modular, research-oriented open-source framework designed for Retrieval-Augmented Generation (RAG) algorithms, providing comprehensive evaluations and reproducibility of six state-of-the-art RAG models. Key features: - Comprehensive RAG Ecosystem: Supports the entire RAG pipeline from data collection and training to auto-evaluation. - Advanced Algorithm Implementations: Reproduces 6 state-of-the-art RAG algorithms, with an easy-to-extend framework for developing new algorithms. - Interact Mode & Evaluation Mode: Designed for quick algorithm understanding and reproducing paper results. - Fair Comparison Platform: Provides benchmark results for 6 algorithms across 5 task types and 10 datasets. - Efficient Retriever Client: Offers local API for parallel access and caching, with average latency under 1 second. - Versatile Generator Support: Compatible with 70B+ models, VLLM, and quantization techniques. - Flexible Instruction Lab: Customizable instruction templates for various RAG scenarios. Tech stack: Python Why it matters: RAGLAB simplifies the process of researching and developing RAG algorithms, making it easier to compare and evaluate different models, and facilitating the creation of new algorithms and datasets. Repository: fate-ubw/RAGLAB Primary language: Python GitHub stars: 314 GitHub description: [EMNLP 2024: Demo Oral] RAGLAB: A Modular and Research-Oriented Unified Framework for Retrieval-Augmented Generation

### Overview: This repository provides a comprehensive implementation of end-to-end speech recognition using TensorFlow, focusing on CTC, Attention, and Multi-Task Learning (MTL) techniques. It includes pre-processing and training options for various corpora such as TIMIT, LibriSpeech, and CSJ. ### Key features: - End-to-End Speech Recognition Implementation - TensorFlow Backend - CTC, Attention, and MTL Training - Support for TIMIT, LibriSpeech, and CSJ Corpora - Pre-processing based on [this repo](https://github.com/hirofumi0810/asr_preprocessing) ### Tech stack: Python ### Why it matters: This repository serves as a foundational resource for developers interested in implementing end-to-end speech recognition systems using TensorFlow, offering a practical approach to handling complex speech recognition tasks with attention mechanisms and multiple task learning.

### Overview: MAUVE is a Python library designed to measure the gap between neural text and human text, providing a similarity score known as MAUVE. This measure is based on Kullback–Leibler (KL) divergences between text distributions quantized in a large language model's embedding space. MAUVE helps identify differences in quality arising from model sizes and decoding algorithms. ### Key features: - **MAUVE with k-means quantization**: Computes MAUVE using *k*-means clustering. - **Adaptive *k*-means hyperparameter selection**: Automatically optimizes *k*-means hyperparameters for better performance. - **Featurization options**: Supports computation using pre-computed GPT-2 features or raw text using HuggingFace Transformers. - **Modality support**: Can be used for other modalities like images or audio by passing pre-computed feature embeddings. ### Tech stack: Python with PyTorch and HuggingFace Transformers. ### Why it matters: MAUVE provides a quantitative measure to evaluate the quality of neural text generation, helping researchers and practitioners understand the discrepancies between machine-generated and human-written text. This can be crucial for improving text generation models and enhancing their ability to mimic human language effectively. Repository: krishnap25/mauve GitHub stars: 314 GitHub description: Package to compute Mauve, a similarity score between neural text and human text. Install with `pip install mauve-text`. README excerpt: This is a library built on PyTorch and HuggingFace Transformers to

### Overview: This project is a Python library called `llm-reader` that converts webpages into a format suitable for LLMs, enhancing the accuracy of web scraping and text extraction tasks. It is an alternative to Firecrawl and Jina Reader API, offering a more open-source and efficient solution for web scraping and LLM integration. ### Key Features: - Converts webpages into LLM-friendly text. - Improves accuracy for web scraping and text extraction. - Supports image and webpage link extraction. - Fully open-source alternative to Firecrawl and Jina Reader API. ### Tech Stack: Python ### Why it Matters: - Enhances the accuracy of web scraping and text extraction. - Provides a more efficient and open-source alternative to existing libraries. - Suitable for use with various LLMs and web scraping tasks. ### Repository: m92vyas/llm-reader - GitHub Stars: 314 - GitHub Description: Turn Webpage to LLM friendly input text. Similar to Firecrawl and Jina Reader API. Makes RAG, AI web scraping, image & webpage links extraction easy. ### Install: ```python pip install git+https://github.com/m92vyas/llm-reader.git playwright install # to download browser. playwright install-deps ``` ### Import: ```python from url_to_llm_text.get_html_text import get_page_source from url_to_llm_text.get_llm_input_text import get_processed_text ``` ### Get processed LLM input text: ```python url = <

Overview: This repository contains the Pytorch implementation of the baseline model used in the 2019 Baidu Joint Relation Extraction (JRE) competition, achieving an F1 score of 0.75 on the development set. Key features: - Pytorch-based baseline model for JRE - Joint relation extraction - Uses BiLSTM for subject prediction followed by object and predicate prediction - Reused code from the original model by Bojone Tech stack: Pytorch Why it matters: This project provides a practical example of how to implement a joint relation extraction model using Pytorch, showcasing the effectiveness of the baseline approach used in the competition. Repository: zhengyima/kg-baseline-pytorch Primary language: Python GitHub stars: 314 GitHub description: 2019 Baidu's Joint Relation Extraction competition, Pytorch baseline, F1 score of 0.75 on the development set. README excerpt: This repository contains the Pytorch implementation of the baseline model used in the 2019 Baidu Joint Relation Extraction (JRE) competition, achieving an F1 score of 0.75 on the development set.

Overview: HPT - Hyper-Pretrained Transformers is a novel multimodal language model framework developed by HyperGAI, designed for edge devices. It has achieved highly competitive results on various multimodal LLM benchmarks. Key features: - HPT 1.5 Edge: A highly efficient model optimized for edge devices, demonstrating impressive capabilities while being extremely efficient. - HPT 1.5 Air: A state-of-the-art model built with Meta Llama 3, achieving the best results across a wide range of benchmarks. Tech stack: Python Why it matters: HPT models are designed to be highly efficient and capable of real-world understanding and complex reasoning, making them suitable for edge devices and edge computing environments.

### Overview: This repository contains a collection of papers and related works focused on large language models, specifically including papers and works related to the development and evaluation of large language models such as ChatGPT, GPT-3, and LLaMA. The papers cover various aspects of these models, including evaluation, survey, in-context learning, instruction tuning, reinforcement learning for human-in-the-loop (RLHF), pre-training techniques, mixtures of experts, knowledge enhanced models, knowledge distillation, knowledge generation, and reasoning. ### Key Features: - Papers and related works on large language models. - Evaluation, survey, in-context learning, instruction tuning, RLHF, pre-training techniques, mixtures of experts, knowledge enhanced models, knowledge distillation, knowledge generation, and reasoning. - Contributors and maintainers of the repository. - Automation script powered by Auto-Bibfile for easy citation management. ### Tech Stack: TeX ### Why it Matters: This repository serves as a comprehensive resource for researchers and practitioners interested in the development and evaluation of large language models. It provides a platform for sharing and accessing the latest research in this field, facilitating further advancements and practical applications.