Projects

Overview: Perceptron is a software designed to help researchers, students, and programmers analyze and optimize artificial neural networks. Key features: - Flexible design for creating, comparing, and testing neural networks. - Educational tool for understanding ANN parameters and their impact on performance. - Visual tools for free and simple ANN analysis. Tech stack: Python Why it matters: This software provides a low-level view of neural network operations, allowing users to experiment and gain insights into the parameters that influence performance.

### Overview: Paragraph Vectors is a PyTorch implementation of the doc2vec algorithm, designed to efficiently process and analyze large collections of text documents. ### Key Features: - **Sparse Updates**: Only sampled noise words are used in forward and backward passes, making the model sparse and computationally efficient. - **Parallel Training**: Batches of training data are generated in parallel on a CPU while the model is trained on a GPU, significantly speeding up the training process. - **Sparse Representation**: The model learns sparse representations of both paragraphs and individual words, capturing the most relevant information efficiently. ### Tech Stack: The implementation is built using PyTorch, a popular deep learning framework. ### Why It Matters: This model is particularly useful for applications that require efficient and effective processing of large text datasets, such as natural language processing tasks, sentiment analysis, and text classification. Its ability to handle sparse data and parallel training makes it well-suited for real-world applications where computational resources are limited.

### Overview: Hallucination is a significant challenge faced by multimodal large language models (MLLMs), hindering their practical applications that require precise judgment. Existing methods mitigate this issue through either specific training data or external knowledge, leading to additional costs. OPERA, introduced in this repository, is a novel approach grounded in an Over-trust Penalty and Retrospection-Allocation strategy, offering nearly free mitigation of hallucination without requiring additional data, knowledge, or training. ### Key Features: - **Over-trust Penalty**: A penalty term added to the model logits during beam-search decoding to mitigate over-trust issues. - **Retrospection-Allocation**: A rollback strategy that retrospectively identifies and re-allocates token selection if necessary, ensuring a more balanced and accurate representation of the input tokens. ### Tech Stack: Python ### Why it Matters: By alleviating hallucination in MLLMs, OPERA enables more accurate and contextually relevant responses, enhancing the practical utility of these models in real-world applications without the need for additional training or data.

### Overview: Articut 中文斷詞暨詞性標記服務是一款簡單易用且高效能的工具，通過現代白話中文語法規則實現了高精度的中文斷詞和詞性標記。Articut API 提供了多種處理方式，包括單個文本處理和大量文本批量處理，適用於各種文本分析和聊天機器人等應用場景。 ### Key features: - **Articut API Website**: https://api.droidtown.co/ - **Document**: https://api.droidtown.co/ArticutAPI/document/ - **Benchmark**: - **Design Goal**: - Online / Docker: HTTP Request - Docker: MultiProcessing - Docker: WebSocket: - **Benchmark Results**: - **Time**: - Articut API: 0.1252 秒 - Multi-Processing (MP): 0.1206 秒 - WebSocket (WS): 0.0677 秒 - **Handling Speed**: - 1K 文本: 155 秒 - 2K 文本: 306 秒 - 3K 文本: 455 秒 - **Large Text Handling**: - 1K 文本: 8 秒 - 2K 文本: 14 秒 - 3K 文本: 17 秒

### Overview: This repository, **Awesome Attention Heads**, serves as a platform to gather the latest research on different types of LLM's attention heads. It also includes a comprehensive survey based on these fantastic works. ### Key Features: - **Four-Stage Framework**: We propose an innovative four-stage framework to analyze the reasoning process of LLMs, inspired by human cognitive neuroscience. - **Survey on Attention Heads**: This repository provides insights into how attention heads contribute to the reasoning process in LLMs. - **Citation Guide**: For those interested in citing our work, we have provided a BibTeX entry: [CITATION.bib](./CITATION.bib). ### Tech Stack: TeX ### Why It Matters: Understanding the role of attention heads in LLMs can enhance our knowledge of model interpretability and improve the design of more transparent and explainable AI systems.

Overview: 小牛视频翻译是一款支持本地视频翻译、字幕翻译和 YouTube 视频翻译下载的 AI 工具，集成自动语音识别与多语言翻译功能，帮助创作者高效完成视频翻译，适用于视频本地化与视频出海场景。 Key features: - 支持本地视频翻译 - 支持字幕翻译 - 支持 YouTube 视频翻译下载 - 生成全新的翻译视频，保留背景音效并替换为新的翻译语音，实现声音和嘴型的精准同步 Tech stack: Python, TensorFlow, PyTorch, OpenCV, GStreamer, WebRTC Why it matters: 通过 AI 技术，小牛视频翻译能够帮助创作者跨越语言障碍，实现视频在全球范围内的更广泛传播与分享，助力视频本地化与视频出海。

### Overview: Attention Map Diffusers is a set of tools designed to visualize and analyze attention maps in Hugging Face's Diffusers models, focusing on cross-attention and spatial-temporal self-attention. ### Key Features: - **Attention Visualization**: Shows attention maps for different models like FLUX.2 Klein, Z-Image-Turbo, PRXPixel, and others. - **Cross-Attention**: Demonstrates cross-attention between text and image, image and image, and video and text. - **Spatial-Temporal Attention**: Highlights spatial-temporal self-attention in video models. ### Tech Stack: Python ### Why It Matters: Attention Map Diffusers provides a clear and intuitive way to visualize attention mechanisms in Diffusers models, aiding developers in understanding how models process information and make decisions.

### Overview: FlowCraft is a production-grade Go SDK designed for building AI agents with long-term memory, knowledge retrieval, and voice capabilities. It provides a modular toolkit for extensible AI applications, enabling developers to assemble applications using the core packages directly or through the Forge and Anvil examples. ### Key Features: - **`core`**: The single platform module for agent execution, graph, tool, model, message, skills, inference, memory contracts, event bus, telemetry, workspace, sandbox, deployment/resource assembly, runtime, sessions, and delegation contracts. - **`craft`**: Application assembly on top of `core`, including the `craft.yaml` definition, compile-time capabilities, and plugin host (MCP plugin processes as MCP servers, host primitives as MCP tools for plugins). - **`driver`**: Provider adapters built on `core`, supporting adapters for OpenAI, Anthropic, ByteDance, MiniMax, DashScope, and TypeSafe. - **`backends/*`**: Platform-specific implementations, including SQLite checkpoints and long-term memory backend with its eval harness. - **`examples/forge`**: A runnable local workspace demo built on the current stack, featuring native deploy/inference/memory scenario configs, an interactive TUI, scripted tests, and persona simulation. - **`examples/anvil`**: The runnable `craft` example, a minimal host application that assembles a Craft from a `craft.yaml` definition, a capability, and a plugin directory, then walks the lifecycle. - **`tools/releasegate`**: Release automation for changeset validation, release

### Overview: LedgerForge is an open-source double-entry ledger designed for AI-native money movement. It records transactions in exact integer units, provides a verifiable audit chain, and enforces spending limits and approvals for AI agents. ### Key Features: - **Exact Money:** Tracks transactions with precise integer units. - **Audit Chain:** Records history with a verifiable audit trail. - **Spending Limits:** Enforces spending policies on every API key. - **AI Agent Management:** Allows AI agents to use their own keys with limits and approvals. ### Tech Stack: Go ### Why It Matters: LedgerForge ensures the integrity and accuracy of financial transactions, providing a robust framework for AI-native money movement.

Overview: mini-VLA is a minimal, beginner-friendly Vision-Language-Action (VLA) model designed to educate researchers and students about modern robot policies that fuse images, text, and states to generate continuous actions. Key features: - Collects data using an expert policy - Trains a small vision-language diffusion policy on collected data - Generates actions using a diffusion policy - Provides scripts for data collection, training, and testing Tech stack: Python Why it matters: mini-VLA serves as a clear, hackable template for understanding VLA design, making it accessible for beginners, researchers, and students alike.

Overview: Juman++ is a morphological analyser that incorporates semantic plausibility using a recurrent neural network language model. It offers improved accuracy and speed over the original Juman++ version. Key features: - Improved accuracy - Significantly faster analysis speed (>250x) - Incorporates semantic plausibility - Supports C++14 and later compilers Tech stack: C++ Why it matters: Juman++ leverages advanced machine learning techniques to enhance the accuracy and efficiency of morphological analysis, making it a valuable tool for natural language processing applications.

### Overview: LightDock is a versatile framework designed for protein-protein, protein-peptide, and protein-DNA docking based on the GSO algorithm. It allows users to customize scoring functions, utilize local gradient-free minimization, and incorporate residue restraints to focus on specific interaction regions. This framework is highly adaptable and compatible with various operating systems, including macOS, GNU/Linux, and Google Colab. ### Key Features: - **Scoring Function Customization**: Users can define their own scoring functions to enhance docking accuracy. - **Local Minimization**: Utilizes gradient-free minimization techniques for efficient docking. - **Residue Restraints**: Supports the inclusion of residue constraints in both receptor and ligand partners, enhancing specificity. - **Cross-Platform Compatibility**: Works seamlessly across macOS, GNU/Linux, and Google Colab, ensuring broad applicability. ### Tech Stack: - Python ### Why It Matters: LightDock offers a robust and adaptable solution for protein docking, making it a valuable tool in structural biology and drug discovery. Its flexibility and compatibility with various platforms make it accessible to researchers and developers alike, enhancing its utility in both academic and industrial settings. ### Repository: The LightDock repository is hosted on GitHub at <https://github.com/lightdock/lightdock>. It has been starred by 413 users and has been cited in several scientific publications, including: - "LightDock: a new multi-scale approach to protein–protein docking" in Bioinformatics. - "LightDock goes information-driven" in Bioinformatics. - "Integrative Modeling of Membrane-associated Protein

Overview: MagicBrush is a large-scale, manually-annotated dataset for instruction-guided image editing, featuring diverse scenarios such as single-turn, multi-turn, mask-provided, and mask-free editing. It comprises 10K source-image, instruction, and target-image triples, supporting the training and evaluation of image editing models. Key features: - Diverse editing scenarios: single-turn, multi-turn, mask-provided, and mask-free editing. - 10K source-image, instruction, and target-image triples. - Supports training and comprehensive evaluation of image editing models. Tech stack: Python Why it matters: MagicBrush is the first large-scale, manually-annotated dataset for instruction-guided image editing, providing a valuable resource for researchers and practitioners in the field of image editing and instruction-guided image manipulation. Repository: OSU-NLP-Group/MagicBrush GitHub stars: 412 GitHub description: [NeurIPS'23] "MagicBrush: A Manually Annotated Dataset for Instruction-Guided Image Editing".

### Overview: DreamLayer Eval is an open-source benchmarking platform designed for researchers, labs, and developers to evaluate and compare image and video diffusion models. It automates the process of setting up prompts, seeds, metrics, and reproducible run logging, ensuring consistent and reproducible results. ### Key Features: - **Automate Prompts, Seeds, Metrics, and Reproducible Run Logging**: Simplifies the process of setting up and running experiments. - **Prompt-to-Image Alignment**: Measures the alignment between prompts and generated images. - **Image Quality and Aesthetic Quality**: Evaluates the quality and aesthetic appeal of generated images. - **Object-Level Prompt Adherence**: Checks how closely the prompts adhere to the specified objects. - **Temporal Video Consistency**: Ensures the consistency of video generation models. - **Reference-Based Image and Video Similarity Metrics**: Compares generated images and videos with reference images and videos. ### Tech Stack: - **React**: Frontend framework for the user interface. - **Flask**: Backend services for running experiments. - **SQLite**: Database for storing run logs. - **ComfyUI**: Integration for image workflows. ### Why It Matters: DreamLayer Eval streamlines the benchmarking process, enabling researchers and developers to compare model outputs more efficiently and consistently. This tool is essential for evaluating the performance of image and video generation models, making it easier to identify the best models for various applications. ### Repository: TheDesignFounder/DreamLayer-Eval ### GitHub Stars: 412 ### Primary Language: Python ### GitHub Description:

### Overview: This repository contains the implementation of the ICLR2024 paper "PnP Inversion: Boosting Diffusion-based Editing with 3 Lines of Code". The paper introduces a novel technique called "PnP Inversion", which rectifies inversion deviations directly within the source diffusion branch using just three lines of code, while leaving the target diffusion branch unaltered. This approach significantly improves edit fidelity and structure/background preservation compared to previous inversion techniques. ### Key features: - PnP Inversion: A technique for improving diffusion-based image editing by directly rectifying inversion deviations within the source diffusion branch. - PIE-Bench: An editing benchmark featuring 700 images with diverse scenes and editing types, complemented by versatile annotations. - Inference: Supports both inference and evaluation of image editing performance. - Evaluation metrics: Focus on editability and structure/background preservation, demonstrating superior results across eight editing methods compared to five inversion techniques. ### Tech stack: Jupyter Notebook ### Why it matters: PnP Inversion offers a significant improvement in edit fidelity and structure/background preservation, making it a valuable technique for enhancing diffusion-based image editing.

Overview: Huskarl is a deep reinforcement learning framework designed for modularity and fast prototyping, utilizing TensorFlow 2.0 and the `tf.keras` API for conciseness and readability. It supports parallel computation of environment dynamics across multiple CPU cores, making it ideal for speeding up on-policy learning algorithms and computationally intensive environments like physics-based ones. Key features: - Deep Q-Learning Network (DQN) - Multi-step DQN - Double DQN - Dueling Architecture DQN - Advantage Actor-Critic (A2C) - Deep Deterministic Policy Gradient (DDPG) - Prioritized Experience Replay Tech stack: Python Why it matters: Huskarl simplifies the process of implementing and prototyping deep reinforcement learning algorithms, making it accessible and efficient for both researchers and developers. Repository: danaugrs/huskarl Primary language: Python GitHub stars: 412 GitHub description: Deep Reinforcement Learning Framework + Algorithms README excerpt:  # Huskarl [](https://badge.fury.io/py/huskarl) Huskarl is a framework for deep reinforcement learning focused on modularity and fast prototyping. It's built on TensorFlow 2.0 and uses the `tf.keras` API when possible for conciseness and readability. Huskarl makes it easy to parallelize computation of environment dynamics

### Overview: This repository contains a collection of Jupyter Notebook tutorials designed to help developers understand and apply PyTorch to solve common Natural Language Processing (NLP) problems. The notebooks cover various aspects of NLP, including text classification, word embeddings, sequence-to-sequence models, and more. ### Key features: - Train a bag-of-words model to predict the sentiment of IMDB reviews - Explore different pretrained word embeddings - Train an RNN to predict the sentiment of IMDB movie reviews - Generate weight loss articles using a character-level RNN language model - Train an RNN-based Seq2Seq model with attention to translate from English to French - Fine-tune a GPT-2 model to generate weight loss articles ### Tech stack: PyTorch ### Why it matters: These notebooks provide practical examples of how to use PyTorch for NLP tasks, making it easier for developers to understand and implement deep learning models in their projects. The tutorials cover a wide range of NLP techniques, from basic text classification to advanced models like Seq2Seq and GPT-2, ensuring developers can tackle various NLP challenges effectively.

### Overview: This dataset contains millions of news articles scraped from various domains, primarily sourced from OpenSources, NYTimes, and WebHose English News Articles. The primary goal is to train deep learning algorithms for fake news recognition. ### Key Features: - **Domain-Specific Labels**: Articles are labeled according to the domain they originated from. - **Scraped from 1001 Domains**: The dataset includes articles from 1001 domains, with some additional NYTimes and WebHose articles to balance classes. - **CSV Format**: The articles are formatted as a CSV file with various fields including domain, type, URL, content, scraped and inserted dates, title, authors, keywords, meta information, and tags. ### Tech Stack: - **Data Source**: OpenSources, NYTimes, WebHose English News Articles - **Scraping Tool**: Scrapy - **Data Processing**: Newspaper Library ### Why It Matters: This dataset is crucial for researchers and developers aiming to improve the accuracy of fake news recognition algorithms. It provides a large, diverse dataset that can be used to train models effectively. ### Repository: The repository is hosted on GitHub at [several27/FakeNewsCorpus](https://github.com/several27/FakeNewsCorpus). ### GitHub Stars: The repository has been starred by 412 developers. ### Primary Language: The primary language of the dataset is unknown. ### GitHub Description: This is an open source dataset composed of millions of news articles mostly scraped from a curated list of 10

Overview: Stable Flow is an official implementation of "Stable Flow: Vital Layers for Training-Free Image Editing" presented at CVPR 2025. It addresses the limitations of Diffusion Transformers (DiT) by identifying "vital layers" for consistent image editing. Key features: - Automatically identifies "vital layers" within DiT for consistent image editing. - Enables real-image editing using improved image inversion methods for flow models. - Demonstrates controlled stable edits, including non-rigid modifications, object addition, and global scene editing. Tech stack: Python Why it matters: This project is significant for its innovative approach to image editing, leveraging the limitations of existing models to achieve more diverse and controlled results without the need for training.

### Overview: An AI tool designed to prevent the spread of COVID-19 by analyzing social distancing protocols using video surveillance. This project automates the detection of social distancing violations and provides real-time analytics, helping to enforce guidelines and prevent the spread of the virus. ### Key Features: - **Real-time Social Distancing Detection**: Automatically detects the extent to which social distancing protocols are followed in the area. - **Anonymity**: Does not collect any data of a particular person. - **Analytics**: Shows real-time analytics of the area, including the number of people present and their risk level. - **Alerts**: Can alert police in case of considerable violations of social distancing protocols. - **Versatile**: Supports both video surveillance and drone-based monitoring. ### Tech Stack: Python ### Why it Matters: This project is crucial in the fight against the spread of the novel coronavirus (COVID-19) by providing an automated solution to enforce social distancing guidelines. It helps in reducing the risk of infection by ensuring that individuals maintain a safe distance from others. ### Repository: Ank-Cha/Social-Distancing-Analyser-COVID-19 ### GitHub Stars: 411 ### Installation: - Fork the repository and download the code. - Download the required files and place them in the same directory. - Install all dependencies. - Run `social_distancing_analyser.py` or `social_distancing_analyser 2.0.py`. ### Credits: - **YOLOv3**: Used for object detection. Referenced from