Projects

### Overview: 本项目收集了近年来深度学习中关于模型解释性的相关论文，按引用次数排序。这些论文涵盖了从知识解释到视觉解释的多种方法，包括量化知识、解释知识蒸馏、组件贡献、复杂值神经网络等。所有论文的PDF已上传至腾讯微云，部分论文需要上SCIHUB查找。 ### Key Features: - **量化知识蒸馏**：量化知识蒸馏方法用于解释知识。 - **高频成分帮助解释卷积神经网络的泛化**：高频成分解释卷积神经网络的泛化能力。 - **Score-CAM：评分加权视觉解释**：一种基于评分的视觉解释方法。 - **知识一致性**：研究神经网络和知识一致性。 - **解释复杂值神经网络**：保护隐私的可解释复杂值神经网络。 - **解释不确定性**：解释模型的预测不确定性。 - **解释图像识别**：解释图像识别模型。 - **解释CNNs**：解释卷积神经网络。 - **视觉commonsense推理**：从识别到认知的视觉共情推理。 - **注意力机制**：学习视觉解释的注意力机制。 - **解释视觉解释**：解释卷积神经网络的视觉解释。 ### Tech Stack: 使用Python、PyTorch、TensorFlow等技术进行实现。 ### Why it Matters: 这些解释方法对于理解复杂的深度学习模型至关重要，有助于提高模型的透明度和可解释性，从而在高风险决策中减少误解。

### Overview: This repository contains a comprehensive list of papers related to Large Language Model-enhanced Recommender Systems (LLM4RS). It covers various aspects of integrating large language models into recommendation systems, including interactive and explainable models, generative frameworks, privacy-preserving recommendations, and more. ### Key Features: - **Chat-REC**: Towards Interactive and Explainable LLMs-Augmented Recommender System - **GPT4Rec**: A Generative Framework for Personalized Recommendation and User Interests Interpretation - **TALLRec**: An Effective and Efficient Tuning Framework to Align Large Language Model with Recommendation - **Privacy-Preserving Recommender Systems with Synthetic Query Generation using Differentially Private Large Language Models** - **Recommendation as Instruction Following**: A Large Language Model Empowered Recommendation Approach - **A First Look at LLM-Powered Generative News Recommendation** - **Sparks of Artificial General Recommender (AGR)**: Early Experiments with ChatGPT - **Zero-Shot Next-Item Recommendation using Large Pretrained Language Models** - **Do LLMs Understand User Preferences? Evaluating LLMs On User Rating Prediction** - **Leveraging Large Language Models in Conversational Recommender Systems** - **Rethinking the Evaluation for Conversational Recommendation in the Era of Large Language Models** - **PALR: Personalization Aware LLMs for Recommendation** - **Prompt Tuning Large Language Models on Personalized Aspect Extraction for Recommendations** - **A Preliminary Study of ChatGPT on News Recommendation: Personalization, Provider Fair

### Overview: Nano World Model is a minimalist repository designed for advancing world model science. It provides a streamlined approach to training and evaluating video world models using diffusion-forcing techniques. ### Key Features: - **Minimal Dependencies:** The repository includes only the necessary components for training, making it easy to set up and run. - **Unified Pipeline:** All training, validation, and evaluation processes are managed using Hydra-based configuration systems, ensuring a clean and organized setup. - **Scientific Transparency:** The codebase is clean and open-source, including model checkpoints, making it easy to replicate and compare results. - **Diverse Applications:** The repository supports long-horizon rollouts, conversion to 3D point clouds, and planning capabilities. ### Tech Stack: The repository uses Python, PyTorch, and CUDA for its operations. ### Why It Matters: Nano World Model offers a straightforward and efficient way to train and evaluate video world models, making it accessible for researchers and developers interested in advancing world model science.

Overview: The repository contains the code developed for the paper "Deep Reinforcement Learning For Sequence to Sequence Models" by Yaser Keneshloo, Tian Shi, Naren Ramakrishnan, and Chandan K. Reddy. This project aims to combine concepts from Reinforcement Learning to improve sequence-to-sequence models, addressing exposure bias and inconsistency between train/test measurements. Key features: The repository includes the source code for implementing seq2seq models using reinforcement learning techniques. It also provides insights into the problems of existing seq2seq models and how reinforcement learning can improve them. Tech stack: Python Why it matters: This project offers a fresh approach to solving complex sequence-to-sequence tasks by integrating reinforcement learning methods with deep neural networks, potentially leading to more effective and robust models.

Overview: RAG-FiT is a framework designed to enhance the performance of LLMs by fine-tuning them on RAG-augmented datasets. It facilitates the creation, training, inference, and evaluation of these datasets, making it a valuable tool for researchers and developers. Key features: - Efficient training using parameter-efficient fine-tuning (PEFT) - Customizable workflows through configuration files - Support for various RAG-specific metrics - Modular design for easy integration and customization Tech stack: Python Why it matters: RAG-FiT simplifies the process of fine-tuning LLMs for RAG tasks, making it easier to experiment with different RAG settings and configurations. It provides a consistent, model-independent format for data persistence and evaluation, making it a versatile tool for researchers and developers working with large language models. Repository: IntelLabs/RAG-FiT GitHub stars: 766 GitHub description: Framework for enhancing LLMs for RAG tasks using fine-tuning. README excerpt: [!WARNING] This project is archived. Intel will not provide or guarantee development, support, or updates. [!CAUTION] This project has known security issues. Comments, suggestions, issues, and pull-requests are welcome. <div align="center"> <img src="assets/rag_fit.png" width="400"/> </div> #### Installation ```sh pip install -e . ``` Optional packages can be installed: ```sh pip install -e .[haystack] pip install -e .[deepeval

Overview: Laya is an open-source AI notification command center that aggregates notifications from Slack, Gmail, GitHub, Jira, Notion, Outlook, and Calendar, powered by local LLMs via Ollama and LM Studio or cloud models like Claude and GPT with user-defined API keys. It supports workflows for complex tasks and provides Action Cards for approval. Key features: - Multi-persona brain: Routes events to specialized AI personas. - Card Workspaces: Agent workflows for complex tasks. - Card Research: Launch on-demand deep research sessions. - Agent inference backends: Run classification/synthesis pipeline. - Spaces: User-defined contexts. - Context Association: Link related cards across platforms. - Cross-platform memory: Entity resolution. - Daily Briefing: Morning summary. - Analytics Dashboard: Track events and costs. - Budget Tracking: Monitor LLM costs. - Chat sidebar: Ask questions and manage rules. - Hybrid search: Combine local and lexical retrieval. - Coherence: Cross-platform entity search. - Corpus: Accept PDF, DOCX, Markdown, text, and CSV files. Tech stack: Python. Why it matters: Laya offers a comprehensive solution for professional orchestration, providing a unified platform for managing notifications across multiple tools, while also offering advanced features like multi-persona brain, card workspaces, and agent inference backends.

### Overview: Local RAG is an open-source tool designed to retrieve augmented generation (RAG) with local Ollama models, allowing for offline access to large language models without relying on third-party services or sensitive data. ### Key Features: - **Local Ollama Chat Models**: Enables seamless interaction with local chat models. - **Ollama or Local Hugging Face Embedding Models**: Supports various models for enhanced retrieval. - **Multiple Ingestion Sources**: Ingests files, GitHub repositories, and websites for diverse data retrieval. - **Streaming RAG Responses**: Provides real-time responses through LlamaIndex. - **Browser-Local Settings Persistence**: Ensures persistent chat history and settings. - **Upload, URL, Repository, and Ingest Guardrails**: Simplifies the process of adding new sources. ### Tech Stack: Python ### Why It Matters: Local RAG offers a practical solution for developers and researchers who need to access large language models without the need for external services or sensitive data. This tool simplifies the process of integrating LLMs into local environments, making it easier to experiment with and deploy models locally.

Overview: This open-source Meme Search Engine and Finder is built to self-host locally using Python, Ruby, and Docker, making it free and accessible for developers to use and integrate into their projects. Key features: - Semantic and keyword search for indexing and retrieving memes. - Local processing for image-to-text extraction, vector embedding, and search. - Documentation for a Search API v1, a CLI, and an experimental Chromium popup integration. - Support for token-authenticated Search API v1, dependency-free local CLI, and an experimental unpacked Chromium popup. Tech stack: Ruby Why it matters: This repository provides the services and web app for indexing, searching, and retrieving memes with semantic and keyword search, making it a valuable tool for meme enthusiasts and developers looking to integrate it into their projects. Repository: meme-search/meme-search GitHub stars: 762 GitHub description: The open source Meme Search Engine and Finder.

Overview: Mordecai is a Python library designed for extracting place names from English-language text and resolving them to their corresponding coordinates and structured geographic information. It leverages Elasticsearch and Geonames to achieve this, providing a powerful tool for geoparsing. Key features: - Geoparsing of place names from text - Resolution of place names to coordinates and structured geographic information - Compatible with Elasticsearch for indexing and searching - Supports Python 3 and spaCy for named entity recognition Tech stack: Mordecai uses Python, Elasticsearch, spaCy, and Keras. Why it matters: Mordecai is essential for anyone working with geospatial data in English-language contexts, providing a robust solution for extracting and resolving place names from text, which is crucial for applications like event geocoding, location-based services, and more. Repository: openeventdata/mordecai Primary language: Python GitHub stars: 765 GitHub description: Full text geoparsing as a Python library GitHub image: paper/mordecai_geoparsing.png GitHub link: https://github.com/openeventdata/mordecai

### Overview: Sno Station transforms your existing agents into a cohesive squad, enhancing productivity and learning through continuous collaboration and dream-like memory consolidation. ### Key Features: - **Memory Sharing**: Agents can pass work to each other and review each other's progress. - **Nightly Dreaming**: Agents dream every night to improve their performance. - **Self-Learning**: Sno Station analyzes your sessions to propose improvements to your agents' skills. - **Continuous Learning**: Agents can pick up where they left off if another agent is interrupted. - **REM Memory Consolidation**: Sno Station strengthens memory across all sessions, enhancing learning and performance. ### Tech Stack: TypeScript ### Why It Matters: Sno Station turns your existing agents into a squad, enhancing productivity and learning through continuous collaboration and dream-like memory consolidation. It runs on your computer, eliminating the need for a server or account, making it a self-contained solution.

### Overview: This repository is a collection of papers accepted at top computer vision conferences, designed to help developers find relevant research quickly. It provides a convenient way to search papers based on specific topics or across multiple conferences. ### Key features: - Papers from AAAI, ACCV, ACM MM, BMVC, CVPR, ECCV, ICCV, ICML, and IJCAI conferences. - Papers are sourced from official websites to ensure accuracy. - Users can search papers by conference, topic, or across multiple conferences. - Papers are organized in a structured format for easy navigation. ### Tech stack: Python ### Why it matters: This repository serves as a valuable resource for developers looking to stay updated with the latest advancements in computer vision research. It provides a comprehensive collection of papers from top conferences, making it easier to find relevant research and stay informed about the latest trends and techniques in the field.

Overview: This repository contains a collection of text deduplication scripts that are ready to use, or modify based on your needs. It includes Near-Duplicate Detection using MinHash + MinHashLSH, SimHash, SuffixArray Substring Exact Deduplication, and Bloom Filter Exact Deduplication. Key features: - MinHash + MinHashLSH for near-duplicate detection - 64 or 128 bit SimHash - SuffixArray Substring exact deduplication - Bloom Filter exact deduplication Tech stack: Python Why it matters: This repository provides a comprehensive set of text deduplication tools that can be easily integrated into various applications, ensuring efficient and accurate data management.

### Overview: This repository contains a curated list of papers related to multimodal and large language models, focusing on the areas of artificial intelligence, computation and language, computer vision, and machine learning. The papers are selected based on their relevance and impact on the field, with a particular emphasis on contributions that offer unique insights and substantial contributions. ### Key Features: - **Survey Papers:** Comprehensive overviews of various topics within multimodal and large language models. - **Position Papers:** Papers discussing the positioning and importance of research in the field. - **Structure:** Detailed structure and planning for future research in the field. - **Reasoning:** Papers focusing on the reasoning and analysis of models. - **Generation:** Papers exploring the generation capabilities of models. - **Representation Learning:** Papers on the representation learning techniques used in models. - **LLM Analysis:** Papers analyzing the performance and capabilities of large language models. - **LLM Safety:** Papers discussing the safety and ethical considerations of large language models. - **LLM Evaluation:** Papers evaluating the performance and accuracy of large language models. - **LLM Reasoning:** Papers on the reasoning mechanisms behind large language models. - **LLM Application:** Papers exploring the practical applications of large language models. - **LLM with Memory:** Papers on models that incorporate memory capabilities. - **LLM with Human:** Papers on the integration of human input in large language models. - **Inference-time Scaling (via RL):** Papers on scaling models for inference time. - **Long-Context LLM:** Papers on models that handle long context inputs. -

### Overview: Recall is a project memory tool designed for Claude Code users who want to save their session summaries locally. It ensures that Claude Code, the AI, is the only AI in the loop, eliminating the need for external models or API calls, thereby keeping costs and usage transparent. ### Key Features: - **Local Memory:** Recall keeps a local log of your sessions, condensing them into a resume-ready summary entirely offline. - **Free Subscription:** It solves the cold-start problem by re-explaining the project each session without additional costs. - **Zero Tokens:** The summary is built locally, so capturing and updating your memory costs nothing beyond your subscription. - **Persistent Memory:** It saves your usage credits by condensing the project summary locally, reducing the number of tokens spent per session. - **Privacy:** It ensures that your transcripts and other sensitive information are never sent to any external model or API, maintaining privacy. ### Why Recall Matters: - **Free Subscription:** It solves the cold-start problem without adding to your monthly bill. - **Cost Savings:** It reduces the number of tokens spent per session, stretching your subscription's usage limits. - **Privacy:** It guarantees that your data is never sent to any external model or API, ensuring privacy. - **Zero-Friction:** It starts working the moment the plugin loads, without requiring any installation or configuration. ### Tech Stack: - Python ### GitHub Stars: 764

Overview: 这个项目是基于PaddlePaddle实现的语音识别系统，专注于中文语音识别，支持Windows和Linux环境下的训练和预测，同时兼容Nvidia Jetson开发板进行推理。项目包含详细的文档教程和多种模型，如DeepSpeech2、Conformer、Squeezeformer等，以及支持的WenetSpeech数据集。 Key features: - 支持中文语音识别 - 支持Windows和Linux环境下的训练和预测 - 支持Nvidia Jetson开发板进行推理预测 - 提供GUI界面部署和Web部署选项 Tech stack: Python Why it matters: 该项目提供了一种高效的语音识别解决方案，特别适用于需要中文识别的场景，如语音转文本应用，具有良好的识别效果和稳定性。 Repository: yeyupiaoling/PaddlePaddle-DeepSpeech Primary language: Python GitHub stars: 764 GitHub description: 基于PaddlePaddle实现的语音识别，中文语音识别。项目完善，识别效果好。支持Windows，Linux下训练和预测，支持Nvidia Jetson开发板预测。

### Overview: Trace is a platform for end-to-end generative optimization of AI agents, built on PyTorch. It enables training AI systems with general feedback, such as numerical rewards or losses, natural language text, and compiler errors. This library provides a fully functional implementation of the platform described in the paper "Trace: End-to-End Generative Optimization for AI Agents," which was accepted to NeurIPS 2024. ### Key Features: - **Trace Node**: Defines a node in the computation graph. - **Bundle Function**: Represents a function that can be optimized. - **PyTorch Backend**: Utilizes PyTorch for efficient computation and training. ### Tech Stack: - Python 3.9+ - Git Large File Storage (for large files) - LiteLLM as the default backend for LLMs - AutoGen for compatibility with AutoGen backend ### Why It Matters: Trace simplifies the process of training AI systems by enabling end-to-end optimization with general feedback, making it easier to train AI agents that can adapt to various forms of feedback. This platform is particularly useful in fields such as scientific computing, parallel programming, and optimization of complex systems. ### Repository: - **Repository**: microsoft/Trace - **GitHub Stars**: 762 - **GitHub Description**: End-to-end Generative Optimization for AI Agents - **GitHub License**: MIT License - **GitHub Tags**: trace-opt[autogen] ### Setup: - **Install**: `pip install trace-opt` - **Development**: `pip install -e .

### Overview: This GitHub repository serves as a comprehensive collection of studies related to the text-to-image (T2I) generation task. It categorizes papers and resources based on various research directions, published years, and conferences to provide a structured overview of the field. ### Key Features: - **Research Directions**: Papers are organized by different research directions such as image synthesis, video editing, and more. - **Publication Years**: Papers are grouped by their publication years to highlight trends and advancements over time. - **Conferences**: Papers are categorized by the conferences they were published in, allowing for easy navigation by conference. - **Topics**: A dedicated section summarizes papers related to specific properties of diffusion models, such as diffusion Transformer-based methods and their integration with LLMs. ### Tech Stack: The primary language used in this repository is TeX, which is commonly used for typesetting documents. ### Why It Matters: This repository is valuable for researchers and developers interested in the field of text-to-image generation, as it provides a centralized resource for understanding the latest advancements and methodologies in the field.

### Overview: ToolOrchestra is an end-to-end RL training framework designed to orchestrate tools and agentic workflows. It surpasses GPT-5 while being much more efficient, using both tools and specialized models to solve complex tasks. ### Key Features: - **Orchestrator-8B**: A state-of-the-art 8B parameter orchestrator model designed to solve complex, multi-turn agentic tasks by coordinating a diverse set of expert models and tools. - **Efficient Training**: Jointly optimized by outcome, efficiency, and preference rewards via end-to-end reinforcement learning. - **Tool Set**: Interacts with a diverse tool set, including basic tools, specialized LLMs, and generalist LLMs. - **GAIA Benchmark**: Ranks #1 on the GAIA benchmark. ### Tech Stack: Python ### Why It Matters: ToolOrchestra leverages efficient RL training to coordinate tools and models, making it a powerful tool for solving complex, multi-turn agentic tasks.

### Overview: The Scale-Recurrent Network (SRN) for Deep Image Deblurring is a state-of-the-art method designed to restore blurry images from low-resolution (LR) versions. This project provides a comprehensive solution for image deblurring using a scale-recurrent network, which is trained on real-world datasets to achieve high-quality restored images. ### Key Features: - **Scale-Recurrent Network (SRN)**: A deep learning architecture that incorporates scale-recurrent structure to handle varying blurring levels effectively. - **Real-World Results**: Demonstrated on real data, showcasing improved PSNR and SSIM metrics compared to previous methods. - **Testing Dataset**: Provides detailed results on both real and synthetic datasets, highlighting the model's robustness. - **Prerequisites**: Requires Python 2.7, Scipy, Scikit-image, numpy, and TensorFlow 1.4 with GPU support. ### Tech Stack: - **Python**: The primary programming language used for implementing the model. - **TensorFlow 1.4**: Utilized for training the deep learning model. ### Why It Matters: This project is significant because it offers a practical solution for image deblurring, which is crucial in various applications such as surveillance, security, and scientific imaging. The model's ability to handle varying blurring levels and produce high-quality restored images makes it highly valuable in real-world scenarios. ### Repository: jiangsutx/SRN-Deblur - **GitHub Stars**: 762 - **GitHub Description**: Repository for Scale-recurrent Network for Deep Image De

### Overview: Transcribro is a private and on-device speech recognition keyboard and service specifically designed for Android. It utilizes Whisper.cpp to run OpenAI Whisper models and Silero VAD for voice activity detection, offering a voice input keyboard for typing with speech. Transcribro supports English and is available on the Accrescent app store for enhanced security. It can be verified using `apksigner` or `AppVerifier` for APK downloads, and is also accessible on a Bluesky post for community engagement and testing. ### Key Features: - **Private and On-Device Speech Recognition**: Recognizes speech input for typing on the device. - **Voice Input Keyboard**: Enables typing with speech. - **Voice Activity Detection (VAD)**: Silero VAD for accurate voice activity detection. ### Tech Stack: Kotlin ### Why It Matters: Transcribro simplifies the process of converting speech to text on the device, providing a seamless typing experience without the need for external apps. Its private and on-device nature ensures privacy and security, making it ideal for sensitive applications or environments where external speech recognition is not acceptable. ### Repository: soupslurpr/Transcribro ### GitHub Stars: 761 ### GitHub Description: Private and on-device speech recognition keyboard and service for Android.