Projects

Overview: LLM-Training is a distributed training framework for large language models powered by Lightning. Key features: - Supports Pre-training, Instruction Tuning, DPO, and ORPO. - Supports dynamic data sampling and selective activation checkpointing. - Compatible with HuggingFace models and alternative implementations for specific architectures. Tech stack: Python Why it matters: LLM-Training provides a robust framework for distributed training of large language models, ensuring efficient resource utilization and flexibility in training methods. Repository: ShinoharaHare/LLM-Training Primary language: Python GitHub stars: 24 GitHub description: A distributed training framework for large language models powered by Lightning.

Overview: AdaExplore is an open-source research codebase designed for LLM-driven GPU kernel engineering. It addresses kernel generation with a two-stage workflow: the Adapt stage improves correctness by turning compile/runtime failures into reusable skill memory, and the Explore stage optimizes each target kernel with a diversity-preserving search. Key features: - Adapts kernel synthesis tasks to improve correctness through failure-driven skill acquisition. - Explores candidate kernels using a diversity-preserving tree search to balance local and structural optimization. - Includes an evaluation harness, synthetic task generation pipeline, and a remote evaluation service for multi-GPU setups. Tech stack: Python Why it matters: AdaExplore provides a comprehensive framework for LLM-driven kernel engineering, addressing the challenges of kernel generation and optimization through its two-stage approach and various features. Repository: StigLidu/AdaExplore GitHub stars: 24 GitHub description: [COLM-LLA 2026] The official implementation for paper "AdaExplore: Failure-Driven Adaptation and Diversity-Preserving Search for Efficient Kernel Generation"

Overview: LLM-Wiki 是一个用于整理和维护大语言模型（LLM）、代理（Agent）、检索与反馈（RAG）以及模型训练与评估（SFT、RLHF）等技术领域的知识图谱。它通过 Obsidian 知识库来实现，旨在帮助用户系统性地学习、准备面试、复盘项目、进行论文精读以及进行长期知识复利。 Key features: - 知识库结构化整理 - 原始素材与提炼知识的分离 - 从素材到知识的提炼过程 - 用于个人知识管理的 AI 协作型知识库 Tech stack: Obsidian 知识库，Markdown，HTML Why it matters: 本项目通过将学习、面试准备、论文精读、工程实践和长期知识复利等材料统一整理，提供了一个系统化、结构化和可复用的知识库。它帮助用户在学习、准备面试、项目复盘、论文精读和工程实践等方面获得更高效和系统化的帮助。同时，通过 AI 协作维护，确保知识的持续更新和维护，为用户提供长期的知识复利。 Repository: VectorPeak/LLM-Wiki GitHub stars: 24 GitHub description: A structured knowledge base for large language model technologies, covering LLMs, Agents, RAG, model training, evaluation methodologies, and AI engineering practices.

Overview: IntAttention is an official codebase for the MLSys 2026 paper "IntAttention: A Fully Integer Attention Pipeline for Efficient Edge Inference". It enables high-fidelity and high-speed language and vision model deployment on ARM CPUs. Key features: - IntAttention replaces the floating-point softmax island with an integer probability path. - It reduces data type conversion overhead, memory traffic, and CPU latency while maintaining competitive accuracy. - It achieves up to 3.7x speedup over FP16 attention and up to 2.0x speedup over conventional INT8 attention pipelines. Tech stack: Python Why it matters: IntAttention is a critical optimization for efficient inference on edge devices, addressing the bottleneck of the probability path in quantized attention pipelines.

### Overview: This repository contains the official implementation and dataset for the paper "Event-Free Moving Object Segmentation from Moving Ego Vehicle" by Zhuyun Zhou et al., presented at IROS'24. The dataset, named DSEC-MOS, is a large-scale collection of moving object segmentation sequences from moving ego vehicles, providing rich motion cues for better video understanding. ### Key features: - **Event-Camera-Based Segmentation**: The repository includes a novel network called EmoFormer that leverages event data for segmentation, distinguishing genuinely moving objects from static backgrounds. - **Large-Scale Dataset**: DSEC-MOS contains 16 sequences with 13,314 frames, divided into 11 training sequences (10,495 frames) and 5 test sequences (2,819 frames). - **Comprehensive Evaluation**: Various mainstream methods are evaluated on the DSEC-MOS dataset to benchmark the performance of different segmentation techniques. ### Tech stack: The repository uses Python for implementation and data processing. ### Why it matters: The DSEC-MOS dataset and EmoFormer network provide a valuable resource for researchers and practitioners in autonomous driving, enabling more accurate and efficient object segmentation from dynamic scenes captured by moving vehicles.

Overview: Titus 2 is a Python implementation of Portable Format for Analytics (PFA), designed for model development and production readiness. It supports Python 3.4-3.9 and includes tools for model producers and PFA manipulation. Key features: - Model producers - PFA manipulation tools - Runtime execution (scoring) engine Tech stack: Python 3.4-3.9 Why it matters: Titus 2 simplifies the transition from statistical model development to large-scale and online production, making it easier to integrate models into existing IT infrastructures. Repository: animator/titus2 Primary language: Python GitHub stars: 24 GitHub description: Titus 2 : Portable Format for Analytics (PFA) implementation for Python 3.4+ GitHub license: MIT License GitHub issues: https://github.com/animator/titus2/issues GitHub repository: https://github.com/animator/titus2 PyPI package: titus2 PyPI version: 1.2.1 PyPI download count: 24 PyPI license: MIT License PyPI project page: https://pypi.org/project/titus2 PyPI version badge: https://badge.fury.io/py/titus2.svg PyPI download count badge: https://pypi.org/project/titus2 GitHub stars badge: https://github.com/animator/titus2/blob/master/README.md GitHub issues badge: https://github.com/animator/titus2/issues GitHub repository badge: https://github.com

Overview: ViPFormer is an efficient vision-and-pointcloud transformer designed for unsupervised point cloud understanding. Created by Hongyu Sun, Yongcai Wang, Xudong Cai, Xuewei Bai, and Deying Li, it surpasses state-of-the-art methods in accuracy, model complexity, and runtime latency. Key features: - ViPFormer is a lightweight vision-and-pointcloud transformer that unifies image and point cloud processing. - It optimizes intra-modal and cross-modal contrastive objectives to learn in an unsupervised manner. - Pretrained ViPFormer models can be transferred to various downstream tasks, including 3D shape classification and semantic segmentation. Tech stack: Jupyter Notebook Why it matters: ViPFormer offers a more efficient and accurate alternative to traditional unsupervised methods, making it suitable for real-world applications requiring fast and effective point cloud understanding.

### Overview: Employee Productivity GenAI Assistant Example is an innovative code sample designed to enhance writing tasks efficiency using AWS serverless technologies and Amazon Bedrock's generative AI models. This example showcases how to integrate AI into writing tasks, providing a streamlined interface and comprehensive tools to streamline the generation of complex documents. ### Key Features: - **Real-time Data Streaming:** Users can interact directly with AI models, as demonstrated by generating a poem about New York. - **Prompt Engineering and Reuse:** Users can create and reuse prompts to engineer templates for various tasks. - **Template Application in Activity:** Templates can be applied to generate specific outputs, such as naming a noise-canceling headphone. - **Activity History Feature:** Logs all interactions, ensuring user activity can be reviewed and tracked. - **Interactive Chat with Templates:** Enhances the utility of templates through natural language interactions. - **Multi-Modality Input:** Supports the upload and use of up to six images, providing context-aware responses. - **Markdown Output Formatting:** Enhances readability and structure of AI-generated content, allowing for better organization of information. ### Tech Stack: This example is built on AWS serverless technologies such as Lambda, API Gateway, DynamoDB, and S3, ensuring scalability, high availability, and security through Cognito. The core of its generative abilities is derived from the powerful AI models available in Amazon Bedrock, delivering tailored and high-quality content swiftly. ### Why It Matters: This GitHub repository serves as a guide for deploying this solution on user-owned AWS accounts, ensuring that anyone can set up and

### Overview: This repository contains my personal notes on research papers from various conferences and journals, focusing on human pose estimation and motion capture techniques. The notes are structured to reflect my understanding of each paper, including the research problem, main contributions, and experimental details. ### Key Features: - **Research Papers**: Notes on papers published in conferences and journals. - **Human Pose Estimation**: Focus on techniques for estimating human poses using sparse IMUs. - **Motion Capture**: Notes on lightweight motion capture methods using smartwatches and head-mounted cameras. ### Tech Stack: - Markdown files for organizing notes. - GitHub for hosting and version control. ### Why It Matters: These notes are valuable resources for anyone interested in human pose estimation and motion capture techniques, providing a structured way to understand the methodologies and findings presented in academic papers. ### Repository: Repository: bryanbocao/awesome-open-papernotes GitHub Stars: 24 GitHub Description: Yet another Ph.D. adventure. ### Primary Language: Unknown ### GitHub Description: Yet another Ph.D. adventure.

Overview: Context Lens is a semantic search knowledge base designed for MCP-enabled AI assistants, transforming any content into a searchable knowledge base. Built on LanceDB vector storage, it brings semantic search capabilities to conversations, indexing local files or GitHub repositories, and works seamlessly with Claude Desktop, Cursor, and other MCP clients. Key features: - Semantic Search - Zero Setup - Serverless Storage - 100% Local & Private - Local & GitHub Indexing - Smart Parsing Tech stack: Python Why it matters: Context Lens transforms traditional keyword search into a more comprehensive and context-aware search method, making it easier for AI assistants to understand and answer questions about the content of any file type, including codebases, documentation, and text files. Repository: cornelcroi/context-lens Primary language: Python GitHub stars: 24 GitHub description: Semantic search knowledge base for MCP-enabled AI assistants. Index local files or GitHub repos, query with natural language. Built on LanceDB vector storage. Works with Claude Desktop, Cursor, and other MCP clients.

Overview: Granular is a fast and scalable dataset format that supports efficient random access and resumable appends. It is designed for high-performance data loading and comes with a high-performance data loader. Key features: - 🚀 **Performance:** High read and write throughput locally and on Cloud. - 🔎 **Seeking:** Fast random access from disk by datapoint index. - 🤸 **Flexibility:** User provides encoders and decoders; examples available. - 👥 **Sharding:** Store datasets into shards to split processing workloads. - 🔄 **Determinism:** Deterministic and resumable global shuffling per epoch. - ✅ **Correctness:** A suite of unit tests with high code coverage. Tech stack: Python Why it matters: Granular is useful for applications that require fast and scalable data access, such as machine learning models, data pipelines, and real-time analytics systems. Its ability to handle large datasets efficiently makes it a valuable tool in various data processing tasks.

Overview: This repository contains 62 unique images categorized into 5 different categories, including chat templates, data types, low-rank adapters (LoRA), quantization, and SFT. The images are freely available for use under the CC BY license. Key features: - 62 unique images - 5 categories: chat templates, data types, LoRA, quantization, and SFT - 60 figures and diagrams of LLMs, quantization, low-rank adapters (LoRA), and chat templates Tech stack: Unknown Why it matters: These images can be used freely in blog posts, slides, presentations, or papers under the CC BY license, making them a valuable resource for developers and researchers in the field of language models and AI.

### Overview: MOSS-Image for Stable Diffusion WebUI Forge Neo is a text-to-image extension that leverages the MOSS-Image SFT model to generate high-quality images with sharp text in English and Chinese. This tool integrates seamlessly with Forge Neo, allowing users to experiment with various image-to-image and layout editing features. ### Key Features: - **Text-to-Image with Sharp Text**: Generate images with clear, sharp text in both English and Chinese. - **Image-to-Image**: Convert images into text-based descriptions. - **Layouts with Boxes**: Place layouts using boxes in the prompt, enhancing the visual structure of the generated image. - **Supports INT4 Checkpoints**: Utilizes cleaner INT4 checkpoints for improved model performance. - **Dynamic Prompts**: Works with the sd-dynamic-prompts extension for more dynamic and precise image generation. ### Tech Stack: Python ### Why It Matters: This extension enhances the text-to-image capabilities of Forge Neo, providing users with more versatile and visually appealing image generation options. The integration with INT4 checkpoints ensures better model performance, and the ability to work with dynamic prompts allows for more dynamic and precise image generation, making it a valuable tool for creators and enthusiasts alike. ### Repository: - **Repository**: eduardoabreu81/moss-image-forge-neo - **GitHub Stars**: 23 - **GitHub Description**: MOSS-Image for Stable Diffusion WebUI Forge Neo: text-to-image with sharp English and Chinese text, img2img and layouts by boxes, with the SFT

Overview: Spring Data Reindexer is a Spring Data module designed for the Reindexer database, providing essential features such as repositories, object mapping, query derivation, type-safe string queries, projections, namespace references, transactions, custom conversions, and Ahead of Time (AOT) optimizations. Key features: - Repositories - Object mapping - Query derivation - Type-safe string queries - Projections - Namespace references - Transactions - Custom conversions - AOT (Ahead of Time) optimizations Tech stack: Java Why it matters: Spring Data Reindexer simplifies the development and management of Spring applications that interact with the Reindexer database, making it easier to implement complex data indexing and querying functionalities.

### Overview: This project aims to validate the performance of cross-request KV prefix caching by reproducing Mooncake's published curve on real traces. The study involves 68,266 requests from 393 Claude Code sessions and 23,608 Mooncake requests through a prefix-cache simulator. The key findings reveal that under capacity pressure, most recomputation comes from tool-calling loops, not from sessions idling past a TTL, and the 5-minute TTL never fires. ### Key Features: - **Prefix-Caching Simulator**: A discrete-event simulator of a cross-request prefix cache. - **Hits are Prefix-Contiguous**: A hit is the longest resident prefix of the block chain, not a set intersection. - **Radix Structure Constrains Eviction**: A block with resident children is not evictable. - **In-Flight Chain Must Be Pinned**: Ensures the in-flight chain is pinned to avoid confusion with dead sessions. ### Tech Stack: - Python ### Why It Matters: Under capacity pressure, most recomputation comes from tool-calling loops, not from sessions idling past a TTL. The 5-minute TTL never fires under capacity pressure, indicating that the policy is more effective than previously believed. ### Repository: - **gauravapiscean/agentic-kv-cache** - **GitHub Stars**: 24 - **GitHub Description**: Reproducing agentic KV-cache policy claims on real traces. 68k requests from 393 Claude Code sessions. LRU is harder to beat than the papers suggest.

**Overview:** UpdAPI is a living benchmark designed to measure how quickly frontier coding agents adapt to real API changes. It tracks and evaluates the performance of AI coding models and agents by analyzing and comparing API evolution over time. **Key features:** - Timestamped, provenance-rich records of API and SDK changes. - Evaluation cases for contemporary coding systems, focusing on the workflow of a coding agent. - A public API Freshness Observatory for continuous benchmarking. **Tech stack:** JavaScript. **Why it matters:** UpdAPI provides a valuable empirical measure of how well AI coding systems keep pace with the software ecosystem, helping developers assess the effectiveness of their coding agents and models.

Overview: Programming Generative AI is a comprehensive video course that guides you through the journey from building simple neural networks to creating advanced models like Stable Diffusion. The repository contains the code, slides, and examples from the course. Key features: - Hands-on examples - From VAEs to Stable Diffusion - Text-to-image pipelines - Personalizing large pretrained models Tech stack: Jupyter Notebook Why it matters: This course provides a practical, step-by-step guide to deep generative modeling, making it accessible and engaging for developers looking to learn and apply advanced AI techniques.

Overview: This repository contains the official implementation of the paper "Delving into the Openness of CLIP" presented at ACL 2023. It provides evaluation codes, pre-trained models, and detailed instructions for installation, data preparation, training, and evaluation. Key features: - Evaluation codes for Extensibility and Stability - Pre-trained models for CLIP, CoOp, DeCLIP, SLIP, FILIP, and DeFILIP - Detailed installation and package requirements instructions - Comprehensive documentation for training, evaluation, and reproducibility Tech stack: Python Why it matters: This repository is essential for researchers and developers interested in the openness and extensibility of CLIP, a state-of-the-art visual language model, providing valuable insights and tools for further research and application development.

Overview: This repository contains the official PyTorch implementation and pretrained model weights for the paper "MB2C: Multimodal Bidirectional Cycle Consistency for Learning Robust Visual Neural Representations," which has been accepted for publication in the Conference ACM MM 2024 (Poster). The repository includes datasets for EEGCVPR40 and ThingsEEG, as well as scripts for data preprocessing and training. Key features: - Pretrained model weights for MB2C - Preprocessed EEGCVPR40 and ThingsEEG datasets - Scripts for zero-shot classification and image reconstruction tasks using MB2C Tech stack: Python Why it matters: This repository provides a comprehensive solution for researchers and developers interested in multimodal learning and visual representation learning, offering practical tools and pretrained models for zero-shot tasks and image reconstruction.

Overview: This project turns your Readwise library into a blazing-fast, self-hosted semantic search engine, complete with nightly syncs, vector search API, Prometheus metrics, and a streaming MCP server for LLM clients. Key features: - Vector search API - Prometheus metrics - Streaming MCP server for LLM clients Tech stack: Python Why it matters: This project allows users to easily turn their Readwise library into a powerful search engine, leveraging advanced semantic search capabilities, real-time metrics for monitoring, and a streaming MCP server optimized for large-scale language model clients, making it highly useful for developers and researchers alike. Repository: leonardsellem/readwise-vector-db Primary language: Python GitHub stars: 24 GitHub description: Turn your Readwise library into a blazing-fast, self-hosted semantic search engine – complete with nightly syncs, vector search API, Prometheus metrics, and a streaming MCP server for LLM clients.