Projects

### Overview Phidias is a generative model designed to create 3D content from text, image, and 3D conditions. It leverages diffusion techniques to enhance the quality and generalization ability of 3D generation, making it a powerful tool for designers and artists. ### Key Features - **Reference-Augmented Diffusion**: Phidias uses diffusion models to generate 3D content, leveraging a reference 3D model to guide the generation process. - **Dynamic Conditioning**: The model dynamically modulates the conditioning strength using meta-ControlNet, ensuring better quality and controllability. - **Self-Reference Augmentations**: Phidias incorporates self-reference augmentations to enable self-supervised training with a progressive curriculum, improving generative performance. - **Versatile Applications**: The model can be applied to various 3D content creation tasks, including text-to-3D, image-to-3D, and retrieving 3D references. ### Tech Stack Phidias is built using Python, TensorFlow, and PyTorch. It utilizes xFormers for efficient GPU acceleration and includes dependencies such as xFormers, torch, torchvision, and other necessary libraries. ### Why It Matters Phidias offers a unified framework for 3D generation that combines text, image, and 3D conditions, making it a significant advancement in the field of 3D modeling. Its ability to enhance the quality, generalization, and controllability of 3D generation makes it a valuable tool for designers and artists, potentially revolutionizing the way 3

Overview: Indic-BERT-v1 is a BERT-based multilingual model designed for 11 Indian languages and Indian-English. It is pre-trained on a novel corpus of 9 billion tokens and evaluated on diverse tasks. Key features: - Supports 12 major Indian languages - Pre-trained on 9 billion tokens - Fine-tuned on various evaluation tasks - Available on HuggingFace for easy use Tech stack: Python Why it matters: Indic-BERT-v1 offers a significant improvement in performance and efficiency compared to other multilingual models, making it a valuable tool for NLU tasks in Indian languages. Repository: AI4Bharat/Indic-BERT-v1 GitHub stars: 297

Overview: CHRONOS is a novel retrieval-based approach to Timeline Summarization (TLS) that iteratively poses questions about the topic and retrieved documents to generate chronological summaries. This method constructs an up-to-date dataset for open-domain TLS, surpassing existing public datasets in terms of size and timeline duration. Key features: - Iterative self-questioning for generating timelines - Up-to-date dataset for open-domain Timeline Summarization (Open-TLS) - Comparable results with state-of-the-art methods in closed-domain TLS, with significant improvements in efficiency and scalability Tech stack: Python Why it matters: CHRONOS offers a novel and effective method for open-domain Timeline Summarization, providing a significant improvement in efficiency and scalability compared to existing approaches.

### Overview: LayoutDM is an official implementation of the paper "LayoutDM: Discrete Diffusion Model for Controllable Layout Generation" presented at CVPR 2023. This repository provides a comprehensive framework for generating layouts using a discrete diffusion model, enabling controllable and high-quality layout generation. ### Key Features: - **Controllable Layout Generation**: The model allows for precise control over the layout of elements, making it suitable for applications requiring fine-grained control over the visual structure. - **Discrete Diffusion Model**: Utilizes a discrete diffusion approach to generate layouts, providing a balance between generality and controllability. - **Pretrained Weights and Clustering Weights**: Includes pre-trained weights and clustering weights for efficient and effective layout generation. ### Tech Stack: - Python 3.7 - CUDA 11.3 - PyTorch 1.10 ### Why It Matters: This model offers significant improvements in layout generation compared to traditional methods, making it highly valuable for applications requiring precise control over the visual structure of layouts. The inclusion of pretrained weights and clustering weights further enhances its utility by providing a starting point for efficient and effective layout generation. ### Repository: - **Repository**: CyberAgentAILab/layout-dm - **GitHub Stars**: 297 - **GitHub Description**: "This repository is an official implementation of the paper titled above." ### Setup: - **Requirements**: - Python 3.7 - CUDA 11.3 - PyTorch 1.1

Overview: AI-Flow is an open-source, drag-and-drop interface for building custom AI workflows, connecting models like GPT-4, Stable Diffusion, FLUX, and more with your own API keys. Key features: - Visual Workflow Builder - Real-Time Monitoring - Parallel Processing - Model Management - Import/Export Tech stack: TypeScript Why it matters: AI-Flow simplifies the process of designing, managing, and monitoring AI-driven workflows, making it accessible and efficient for developers.

### Overview: This project, **LRV-Instruction**, introduces a novel method to mitigate hallucination in large multi-modal models. It leverages robust instruction tuning to enhance the performance and reliability of these models. ### Key features: - **Robust Instruction Tuning**: The project uses instruction tuning to improve the accuracy and reliability of large multi-modal models. - **Benchmarking**: It includes a benchmarking tool called **HallusionBench** to evaluate the effectiveness of the model in reasoning tasks. - **Model Support**: The project supports training on V100 GPUs with 32GB of memory. ### Tech stack: Python ### Why it matters: This approach addresses a significant challenge in the field of large multi-modal models by providing a method to mitigate hallucination, ensuring more reliable and accurate model performance.

Overview: BeyondLLM is an all-in-one toolkit designed for rapid experimentation, evaluation, and observability of Retrieval-Augmented Generation (RAG) systems. It simplifies the process by automating integration, providing customizable evaluation metrics, and supports various Large Language Models (LLMs) tailored to specific needs. Key features: - Automated integration of LLMs and embeddings - Customizable evaluation metrics - Support for different LLMs and embeddings - Simplified deployment and observability Tech stack: Python, Jupyter Notebook Why it matters: BeyondLLM aims to reduce LLM hallucination risks and enhance the reliability of Large Language Models (LLMs) by providing a streamlined and efficient way to experiment, evaluate, and observe RAG systems. Repository: aiplanethub/beyondllm GitHub stars: 297 GitHub description: Build, evaluate, and observe LLM apps

### Overview: Nano 是一个基于 Transformer 结构的语言模型，旨在提供个人、研究和应用方面的多种功能。它支持在各类设备上高效推理，包括树莓派、单板机、浏览器、PC 和服务器等。 ### Key features: - **训练实践**：实现 Transformer 语言模型的预训练和监督微调，训练出若干可用的模型。 - **推理实践**：在单板机、单片机、浏览器、PC、服务器等各类环境上，高效推理语言模型。 - **开放探索**：研究 LLM 的动力学等问题，开放探索其他模态和领域的问题。 - **硅碳互动**：对 LLM 祛魅，在实践中建立起对于 LLM 的感性经验、合理预期和哲学认识。 ### Tech stack: C ### Why it matters: Nano 是一个基于 Transformer 结构的语言模型，旨在提供个人、研究和应用方面的多种功能。它支持在各类设备上高效推理，包括树莓派、单板机、浏览器、PC 和服务器等。通过训练实践和推理实践，Nano 为用户提供了一个强大的工具来探索和研究语言模型。

Overview: MagicDrive3D is an open-source project that implements the cutting-edge paper "MagicDrive3D: Controllable 3D Generation for Any-View Rendering in Street Scenes". This project allows developers to create visually stunning 3D street scene images with precise control over lighting and perspective. Key features: - Real-time 3D rendering - Adjustable lighting and shadows - Support for various street scene styles - Easy integration with popular 3D libraries Tech stack: Python, PyTorch, OpenCV Why it matters: MagicDrive3D provides a powerful tool for developers looking to enhance their 3D rendering capabilities, enabling them to create highly detailed and realistic street scenes with just a few lines of code. This project is essential for anyone interested in creating visually stunning and interactive 3D content.

### Overview: The QGate-Sln-MLRun solution is a comprehensive quality gate designed for the MLRun project, ensuring the robustness and reliability of the implementation and delivery. This project includes independent tests for various aspects such as project creation, feature set management, data ingestion, pipeline execution, feature vector generation, model building, serving, monitoring, and performance testing. The solution also provides external and independent coverage, making it essential for ensuring the quality and compatibility of MLRun implementations in real-world environments. ### Key Features: - **Independent Quality Tests**: Comprehensive tests covering all necessary aspects of the MLRun implementation. - **Deeper Quality Checks**: Pre-rollout checks to identify potential compatibility issues. - **External and Independent Coverage**: Ensures thorough testing through external and independent sources. - **Community Support**: Offers community support and resources for better project management and improvement. ### Tech Stack: - Python ### Why It Matters: - Ensures the quality and reliability of MLRun implementations. - Provides a robust quality gate for real-world use cases. - Helps identify and fix potential issues before deployment. - Offers independent and external coverage for comprehensive testing. ### Repository: - **Repository**: george0st/qgate-sln-mlrun - **Stars**: 297 - **Description**: The Quality Gate for solution [MLRun](https://www.mlrun.org/) (and [Iguazio](https://www.iguazio.com/)). The main aims of the project are: - independent quality test (function, integration, performance, vulnerability, acceptance, ... tests

Overview: ShuffleNet is an efficient convolutional neural network architecture designed for mobile devices, achieving better performance than Google's MobileNet. Key features: - Efficient architecture using residual blocks with pointwise group convolutions. - Trained on ImageNet with groups=3 and no channel multiplier. - Pretrained state dictionary available for download. Tech stack: Python Why it matters: ShuffleNet offers a more efficient alternative to MobileNet, potentially improving performance on mobile devices without significant performance degradation.

### Overview: Memex is a powerful tool for indexing, browsing, and revisiting conversations from various coding tools and agents. It supports keyword search, local and remote session management, and token tracking, making it an essential resource for developers and coders. ### Key Features: - **Local Agent History Indexing**: Memex indexes conversations from Claude Code, Codex, Pi, OpenCode, and Github Copilot, allowing users to find previous work and read original transcripts. - **Resume Sessions**: Enables users to resume sessions in their original tool, facilitating seamless continuation of sessions. - **Search and Track Tokens**: Supports searching and tracking token usage from local logs, making it useful for monitoring and managing session data. - **Cross-Tool Sessions**: Allows users to search and import saved sessions from different tools, enhancing collaboration and knowledge sharing. - **Engine Support Table**: Lists supported engines and their capabilities, including history, resume, and token usage. ### Tech Stack: Rust ### Why It Matters: Memex is a valuable tool for developers and coders who need to quickly find and revisit their work, collaborate with multiple tools, and manage session data efficiently. Its ability to index and resume sessions across different tools makes it a versatile and indispensable resource in the developer's toolkit. Repository: nicosuave/memex GitHub Stars: 291

Overview: py-idp is an AI-enabled Intelligent Document Processing library for Python, designed to handle a wide variety of document types efficiently. It includes 23 LLM backends, Pydantic schemas, built-in schema discovery, and a self-hosted OCR using Nanonets-OCR2-3B, among other features. Key features: - Auto-schema discovery - Auto-chunking for oversized documents - 23 LLM backends - Built-in eval harness - Pydantic schemas-driven Tech stack: Python Why it matters: py-idp offers a comprehensive solution for AI-driven Intelligent Document Processing, providing robust features for handling various document types, including advanced schema discovery and automatic chunking for large documents. It also supports multiple LLM providers and a self-hosted OCR, making it versatile and powerful for developers. Repository: rollroyces/py-idp Primary language: Python GitHub stars: 297 GitHub description: AI-enabled Intelligent Document Processing for Python: 12+ LLM backends, Pydantic schemas, self-hosted OCR (Nanonets-OCR2-3B), auto-schema discovery, HITL review, prod FastAPI server. AGPL-3.0 + commercial.

### Overview: Tiny Claw is a self-improving, personal AI companion designed for simplicity, affordability, and ease of use. It aims to bring the personal computer revolution to AI agents, making them accessible and affordable for everyone. ### Key Features: - **Self-Improving Memory**: Smart routing that tiers queries to cut costs. - **Personal Companion Personality**: Unique, self-configurable, and grows with you. - **Native Framework Built from Scratch**: Cheap to run and treats users as friends. - **Smart Routing**: Adapts to whichever provider plugins you install. - **Self-Configuring Setup**: Zero hassle, self-configuring. - **Cheap and Affordable**: Built from scratch, expensive models are tiered, and every query hits the most expensive model. ### Tech Stack: TypeScript ### Why It Matters: Tiny Claw is a personal AI companion that aims to bring the personal computer revolution to AI agents, making them accessible and affordable for everyone. It is built from scratch, cheap to run, and treats users as friends, not just as users to be served.

### Overview: SkillX is a fully automated framework designed to construct a reusable, plug-and-play skill knowledge base for LLM agents from experience. It distills agent experience into a three-level skill hierarchy: Planning Skills, Functional Skills, and Atomic Skills, improving both task success and execution efficiency across challenging benchmarks such as AppWorld, BFCL-v3, and τ2-Bench. ### Key Features: - **Hierarchical Multi-Level Skill Design**: SkillX transforms raw trajectories into a structured three-tier skill space: Planning Skills, Functional Skills, and Atomic Skills. - **Fully Automated Skill KB Construction**: SkillX provides an end-to-end automated pipeline that rolls out agents on training tasks, extracts reusable skills from successful trajectories, consolidates and filters low-quality skills, and builds a reusable skill knowledge base. - **Iterative Skill Refinement**: SkillX continuously improves the skill library through skill merging, quality filtering, and iterative updates based on execution feedback. - **Exploratory Skill Expansion**: SkillX proactively discovers new skills by identifying under-used and failure-prone tools, guiding environment exploration, synthesizing new tasks from exploratory trajectories, and expanding skill coverage beyond the original training distribution. - **Plug-and-Play Transfer Across Agents**: The resulting skill library can be directly injected into different base agents, enabling strong-to-weak transfer without retraining the underlying models. ### Tech Stack: Python ### Why It Matters: SkillX consistently improves both task success and execution efficiency across challenging benchmarks, making it a valuable tool for enhancing the performance and adaptability of LLM agents.

### Overview: Humanizer is a free, open-source AI writing humanizer and detector. It reads like a human wrote it, with 55 patterns and 5 voices. It scores text from 0 to 100 based on four measurable signals, providing a 0-100 AI-tell score. The CLI allows zero setup and nothing leaves your machine. ### Key features: - **Score Text**: Scans your text and scores it from 0 to 100. - **55 Patterns and 5 Voices**: Offers 55 patterns and 5 voices to match your writing style. - **No Setup Required**: No need for any setup or configuration. - **0-100 AI-tell Score**: Scores text from 0 to 100 based on four measurable signals. ### Tech Stack: - JavaScript ### Why it Matters: Humanizer is a powerful tool for improving your writing. It reads like a human wrote it, making your writing more engaging and natural. The 55 patterns and 5 voices allow you to match your writing style, making your work stand out. The zero setup and 0-100 AI-tell score make it easy to use and effective. ### Repository: Repository: Aboudjem/humanizer-skill GitHub Stars: 296 GitHub Description: Free, open-source AI writing humanizer and detector. 55 patterns, 5 voices, a 0-100 AI-tell score, and nothing leaves your machine. ### Primary Language:

### Overview: FireFlow: An 8-Step Inversion and 8-Step Editing Process for Image Semantic Editing FireFlow is an innovative approach that leverages the FLUX-dev model to achieve significant improvements in image semantic editing. By employing an 8-step inversion and 8-step editing process, FireFlow demonstrates a 3x speedup in performance with results that are both comparable and superior to baseline methods. ### Key Features: - **8-step Inversion Process**: FireFlow's inversion process is designed to be highly efficient and scalable, achieving second-order precision at the cost of a first-order method. - **8-step Editing Process**: The editing process is also optimized, providing a comprehensive solution for image semantic editing tasks. - **3x Speedup**: The overall performance of FireFlow is significantly improved compared to previous methods, offering a substantial speedup in tasks such as image reconstruction and semantic editing. ### Tech Stack: Python FireFlow is built using Python, leveraging the powerful libraries and frameworks available in the Python ecosystem. This choice allows for easy integration and scalability of the codebase. ### Why It Matters: FireFlow's ability to achieve significant improvements in image semantic editing tasks, while maintaining performance comparable to or even superior to baseline methods, makes it a valuable tool for researchers and practitioners in the field of computer vision and image processing. The 3x speedup offered by FireFlow is particularly noteworthy, making it an attractive option for applications that require real-time performance or high-throughput tasks. Repository: HolmesShuan/FireFlow-Fast-Inversion-of-Rect

Overview: FireAct is a project that focuses on fine-tuning language agents for various tasks, including hotpotqa. This repository contains prompts, demo code, and fine-tuning data generated based on the publication "FireAct: Toward Language Agent Fine-tuning." Key features: - Fine-tuning of language models for specific tasks - Generation of training data and experiments - Data and prompts for both Alpaca and GPT formats Tech stack: Python Why it matters: FireAct provides a practical approach to fine-tuning language models for agents, making it useful for researchers and developers looking to improve their language models for specific applications.

### Overview: This repository is a curated list of Machine Learning (ML) and Generative AI (GenAI) libraries and resources specifically for the Elixir programming language. It serves as an overview for experienced Elixir developers and can also be useful for ML and AI practitioners looking for alternative ecosystems. ### Key features: - Nx - Tensors for Elixir with compilation to CPU/GPU - Explorer - Series and dataframes for data exploration in Elixir - Livebook - Write interactive and collaborative notebooks, with integrations to databases, messaging, visualization, and more - Kino - Render rich and interactive output - Pythonx - Embeds a Python interpreter directly into Elixir via NIF, running in the same OS process as the BEAM - EMLX - Nx backend and compiler for Apple MLX - Emily - Alternative MLX-based Nx backend with fused transformer kernels - Handoff - Nx-compatible library for building and executing distributed DAGs of dependent functions across BEAM nodes - NxSignal - Digital signal processing on Nx: FFT, filters, and spectrograms - Tokenizers - Bindings to Hugging Face Tokenizers - EXGBoost - Decision Trees implemented using the XGBoost C API - Mockinjay - Implementation of Microsoft's Hummingbird library for converting trained Decision Tree models into Nx tensor computations - Soothsayer - Time series forecasting library inspired by Facebook's Prophet and NeuralProphet - Ulam - Elixir interface to Stan, a probabilist programming language - eXMC - Probabilistic programming on

Overview: HanLP中文分词Lucene插件是基于HanLP的中文分词工具，适用于包括Solr在内的基于Lucene的系统。 Key features: - 支持Solr和其他基于Lucene的系统 - 使用HanLP分词器进行分词 - 支持索引和查询模式 - 提供停用词和同义词过滤功能 - 支持多种分词算法，如Viterbi算法 Tech stack: Java Why it matters: HanLP中文分词插件为开发者提供了强大的中文分词功能，适用于多种基于Lucene的搜索引擎和全文搜索系统，提高了搜索结果的准确性和用户体验。