Projects

### Overview: This repository contains a Keras implementation for the ISIC 2019 Challenge Task 1: classify dermoscopic images among nine different diagnostic categories without meta-data. The approach achieved a Balanced Multiclass Accuracy of 0.505, ranking 25th out of 64 unique teams. The key features include the use of ODIN for out-of-distribution detection and the implementation of the Seven-Point Checklist Dermatology Dataset. ### Key features: - Implementation of the Seven-Point Checklist Dermatology Dataset - Use of ODIN for out-of-distribution detection - Balanced Multiclass Accuracy of 0.505 - Ranking 25th out of 64 unique teams ### Tech stack: Keras with TensorFlow backend ### Why it matters: This project showcases the effectiveness of using out-of-distribution detection methods in medical image classification tasks, particularly in challenging scenarios where meta-data is not available.

Overview: This repository contains a collection of fullstack GenAI applications and AI agents designed to help developers build cutting-edge applications using various technologies. Key features: - Fullstack Multimodal RAG (Revised) - Product Research AI Agent - AI Podcast Generator - Text 2 AI Video Generator - AI Agent for Crypto Market Analysis Tech stack: Jupyter Notebook Why it matters: These applications showcase the versatility and integration of different technologies in creating intelligent and user-friendly AI solutions.

Overview: ContextQA is an open-source tool designed for leveraging the power of Large Language Models (LLMs) and vector databases to facilitate data-driven conversations. Key features: - Supports regular chat with knowledge expansion through internet access. - Implements Conversational QA with relevant sources. - Provides streaming responses. - Enables ingestion of data sources used in QA sessions. - Manages data sources. - Offers settings for LLMs and vector databases, including providers, models, temperatures, and engines. Tech stack: Python Why it matters: ContextQA simplifies the integration of LLMs and vector databases into applications, making it easier to build conversational AI solutions that can access and utilize vast amounts of structured and unstructured data. Repository: zaldivards/ContextQA GitHub stars: 21 Primary language: Python GitHub description: ContextQA - The open-source tool for data-driven conversations

### Overview: Train_llm is a small chat model trained from scratch, utilizing various techniques such as tokenizer, pretraining, SFT, reinforcement learning, and multimodal capabilities. The model is highly efficient with 0.78B parameters and a Core score of 0.248, showcasing its effectiveness across multiple metrics. ### Key Features: - **Tokenizer**: Trained using RustBPE on a large dataset. - **Pretraining**: Fine-tuned on Karpathy's FineWeb-Edu-100B-Shuffle dataset with 8 tokens per parameter. - **SFT (Sequence-to-Sequence Fine-tuning)**: Utilizes nanochat's mixture of SmolTalk, MMLU, and GSM8K. - **Reinforcement Learning (RL)**: Employs a KL penalty and supports tool calling for a calculator and various tools. - **Multimodal Capabilities**: Includes image understanding and conversational image generation. - **Tool Calling**: Supports a calculator and any tool defined in the hermes function-calling format. ### Tech Stack: - **Python**: The primary programming language used for the project. - **Hugging Face**: Utilizes weights on Hugging Face for training and evaluation. ### Why It Matters: - **Efficiency**: The model achieves a Core score of 0.248, which is close to GPT-2's performance at 0.256. - **Versatility**: The model excels in various tasks including CORE, SFT, RL, and multimodal capabilities.

Overview: This project aims to modernize and vectorize cyber operations data by transforming it into embeddings using Qdrant, enabling fast similarity search and context-aware analysis for AI-driven systems. Key features: - Vectorizing cyber operations data - Building a vectorized data pipeline - Mapping external and internal attack surfaces into vectors - Visualizing risky assets using dimensionality reduction and graph-style views Tech stack: Python Why it matters: This approach simplifies the integration of AI-driven analysis with legacy data, making it easier to visualize and understand complex cyber operations data in a way that AI can easily process and analyze.

### Overview: This repository collects and maintains high-quality instruction fine-tuning datasets across various domains and languages. The primary goal is to provide a resource for researchers and developers to train large language models (LLMs) effectively. The datasets are designed to be accessible and useful, even if they require private knowledge datasets due to concerns about data privacy. ### Key Features: - **Instruction Fine-Tuning Datasets**: A collection of high-quality instruction datasets for fine-tuning large language models. - **Cross-Lingual Benchmarking**: Datasets for evaluating cross-lingual generalization abilities of multilingual models. - **Multilingual Knowledge Transfer**: Resources for transferring knowledge from private datasets to instruction datasets, suitable for training LLMs. - **Code Generation**: Datasets for code generation tasks, including CodeSearchNet, CodeXGLUE, ConCode, and code-docstring-corpus. - **General Knowledge**: Resources for general knowledge tasks, such as Huggingface datasets for MQA, MLQA, and XNLI. - **Multilingual Datasets**: Datasets for multilingual instruction fine-tuning, including GuanacoDataset and CMRC2018. ### Tech Stack: Python, Huggingface, and various datasets. ### Why It Matters: This project aims to bridge the gap in high-quality instruction datasets, making it easier for researchers and developers to train LLMs effectively. The availability of these datasets can significantly enhance the performance and capability of LLMs in various domains, including cross-lingual and multilingual instruction fine-tuning.

Overview: Tessmora is a self-hosted multimodal RAG with an autonomous research agent designed for efficient document, image, audio, and video search with traceable citations. Key features: - Unified search across various media types - Retains original text for documents and table headers/rows for tables - Supports finding keywords beyond the text - Integrates Dense, BGE-M3 Sparse, CLIP, and CLAP channels for enhanced recall and reordering - Enables autonomous research with the Pi Agent Mode, allowing the Agent to search, read, extend context, view media, and submit answers with citations - Checks the correctness of the conclusion by referencing original files, text context, images, and video time ranges - Clearly defines materials and search scope - Supports web UI, API, CLI, and Docx/Wiki import Tech stack: Python 3.11 or 3.12, Node.js 22.19 or higher, Docker, Compose, FFmpeg/ffprobe, and LibreOffice Why it matters: Tessmora simplifies the process of searching and analyzing various media types, ensuring accurate and traceable citations, making it an essential tool for researchers and knowledge workers. Repository: Champ-X/MMA-RAG Primary language: Python GitHub stars: 18 GitHub description: Tessmora — self-hosted multimodal RAG with an autonomous research agent. Search documents, images, audio, and video with traceable citations.

Overview: A versatile chatbot UI designed for seamless integration with various models and environments, supporting RAG, multimodal, and text completion functionalities. Key features: - Vanilla chat interface - Chat with short document - Chat with visual image - Chat with visual image and short document - Chat with long document via RAG - Text completion (free form prompting) - Batch inference (via file upload with vLLM) Tech stack: Python Why it matters: This chatbot UI provides a unified interface for accessing a wide range of AI models, making it easier to integrate with different environments and models, including Transformers, llama.cpp, MLX, and vLLM. Repository: DAMO-NLP-SG/Multipurpose-Chatbot GitHub stars: 20 GitHub description: A chatbot UI for RAG, multimodal, text completion. (support Transformers, llama.cpp, MLX, vLLM)

Overview: This repository serves as a comprehensive training and evaluation framework for ECG-Language Models (ELMs), providing a unified way to prepare datasets and evaluate models. It supports multiple architectures, training objectives, and data representations, including distributed training capabilities. Key features: - Supports multiple ELM architectures and training objectives. - Includes a unified data structure for ECG datasets, ensuring compatibility with various models. - Provides a preprocessed ECG dataset repository for easy access. - Supports pretraining of ECG encoders using the ECG-Neural-Networks repository. Tech stack: Python Why it matters: This framework is essential for researchers and developers interested in advancing the field of ECG-Language Models. It simplifies the process of preparing datasets, training models, and evaluating performance, making it a valuable tool for both academic research and practical applications in healthcare and biomedical engineering. Repository: ELM-Research/ECG-Language-Models GitHub stars: 20 GitHub description: A research-oriented training and evaluation framework for ECG-Language Models (ELMs) README excerpt: A Training and Evaluation Framework for ECG-Language Models (ELMs)

Overview: TurboQuant is a pure Python implementation of the TurboQuant algorithm, designed for FAISS-compatible vector quantization. It offers 95% recall with 5x compression, no preprocessing, and no GPU requirements, making it a significant improvement over FAISS. Key features: - **FAISS-compatible**: Built to seamlessly integrate with FAISS for vector similarity search. - **95% Recall**: Achieves high recall rates, ensuring accurate results. - **5x Compression**: Offers significant compression without compromising on recall. - **No Preprocessing**: Eliminates the need for additional preprocessing steps. - **Pure Python**: Utilizes Python for all computations, making it lightweight and easy to integrate. Tech stack: Pure Python, NumPy, and FAISS-compatible. Why it matters: TurboQuant is a game-changer for AI applications that require efficient vector quantization, such as in KV cache compression, where it outperforms traditional methods like FAISS by 5x in terms of compression and 95% recall in terms of recall. Repository: Firmamento-Technologies/TurboQuant GitHub stars: 20 GitHub description: Near-optimal vector quantization from Google's ICLR 2026 paper — 95% recall, 5x compression, zero preprocessing, pure Python FAISS replacement.

### Overview: ArtiControlNet 是一个辅助设计的 AIGC 系统，旨在简化线稿上色、平面设计、绘画参考、传图分析和服装设计等任务。通过描述你的需求，系统将生成相应的图像或提供详细的图像分析。 ### Key features: - **文生图**: 描述你想要的画面，生成一张新图。 - **线稿生图**: 手绘线稿传上来，指定颜色与风格，成品出来。 - **图像问答**: 传图问它，评图、打分、逐像素分析都行。 ### Tech stack: Python 3.10+ | FastAPI | Vue 3 | LangGraph ### Why it matters: ArtiControlNet 提供了高效、便捷的设计辅助工具，帮助设计师节省时间和精力，提高工作效率。其强大的功能和用户友好的界面设计使其成为设计行业的有力工具。

Overview: Aktiva AI is a self-hosted Discord bot designed for managing and interacting with multiple AI personalities on the platform. Key features: - Seamless character swapping - Channel-based memory - Thread support - Image recognition - Character message editing and deletion - Customizable AI characters - PDF file reading support - Web search integration - Whitelist management - OpenRouter API integration - SillyTavern character compatibility Tech stack: Python Why it matters: Aktiva AI offers a unique and engaging way to interact with multiple AI personalities on Discord, providing a personalized and immersive experience.

Overview: 这个项目提供了一个开箱即用的医疗文本二分类微调框架，旨在简化医疗大语言模型的训练过程，提升训练效率和灵活性。 Key features: - 多模式支持：兼容 QLoRA、LoRA 和全量微调。 - 硬件友好：深度集成 DeepSpeed，支持多卡 ZeRO 优化，缓解 OOM 焦虑。 - 配置驱动：将繁琐的超参抽离为 YAML 配置，代码与参数解耦，易于复现。 Tech stack: Python Why it matters: 降低医疗大语言模型的训练门槛，提供高效、灵活的训练体验。 Repository: Jackohhhh/MedLLM-Finetuning GitHub stars: 20 GitHub description: 一个开箱即用、用于二分类任务的大语言微调模型框架。An out-of-the-box LLM fine-tuning framework for medical binary classification.

Overview: SparkOffer 是一个面向技术岗求职者的 AI 面试训练系统，旨在通过持续的个性化能力建模和闭环反馈，帮助求职者提升面试技巧和技能。 Key features: - 个性化出题引擎（三层信息融合） - 训练 → 评估 → 画像更新 闭环 - 多场景训练入口 - 双层知识增强（RAG） - 可观测的成长 Tech stack: FastAPI · LangChain · LangGraph · LlamaIndex Why it matters: SparkOffer 提供了系统化的面试训练方法，通过持续的反馈和能力迭代，帮助求职者在面试中更加自信和高效。

### Overview: This project is an AI-powered CLI designed for file reorganization, allowing users to organize their files locally without the need to upload data to any external servers. ### Key features: - **AI-Powered File Reorganization**: The CLI uses AI models to suggest folder names and organize files efficiently. - **Local Processing**: The app runs entirely on the user's machine, ensuring no data leaves the machine. - **Command Line Interface (CLI)**: Supports multiple commands for easy file organization. - **Process**: The user inputs the source and destination folder paths, AI model name, and embedding model name. - **Destination Folder Scan**: Scans the destination folder for embeddings and compares them to the source folder's files. - **Clustering & AI Folder Naming**: Groups unmatched files using agglomerative hierarchical clustering and suggests folder names to the user. - **Preview Results**: A table displays the proposed destination for each file. - **Apply & Rollback**: Users can apply or rollback changes after the process. ### Tech stack: Rust ### Why it matters: This project leverages AI to streamline file organization, ensuring local processing and efficient file management without the need for external data storage. ### Repository: PerminovEugene/messy-folder-reorganizer-ai ### GitHub stars: 20 ### GitHub description: 🤖 AI-powered CLI for file reorganization. Runs fully locally — no data leaves your machine. ### Build badge: [![codecov](https://codecov.io/gh/PerminovEugene/messy-folder-re

Overview: Libargus is a unified native AI runtime that allows Java applications to run language models, generate embeddings, process images and video, transcribe audio, and synthesize speech directly within the Java application. It supports Java 17+, Java 22+, Panaman FFM, and Android, providing a seamless and efficient way to integrate AI capabilities into Java applications without the need for separate inference services. Key features: - Runs GGUF language models - Generates embeddings - Processes images and video - Transcribes audio - Synthesizes speech - Supports Java 17+, Java 22+, Panaman FFM, and Android Tech stack: Libargus is built on established engines including llama.cpp, GGML, and whisper.cpp, providing a high-performance native inference runtime. Why it matters: Libargus simplifies the integration of AI capabilities into Java applications, eliminating the need for separate inference services, and offers a unified API that integrates seamlessly with existing Java applications, making it a valuable tool for developers looking to enhance their Java applications with AI capabilities.

### Overview: SlideDP is a synchronous data-parallel runtime designed for full-parameter LLM fine-tuning across multiple GPUs, enabling large models to be trained beyond GPU memory limitations. It optimizes parameter delivery, gradient aggregation, and CPU updates around a shared authoritative host state, while decoupling persistent state ownership from communication routes. ### Key features: - **Parameter Delivery**: Synchronous data-parallel runtime for efficient parameter delivery. - **Gradient Aggregation**: Aggregates gradients from multiple GPUs in a single step for optimal communication. - **CPU Updates**: Decouples CPU updates from communication routes, improving resource utilization. - **Workload and Topology Configuration**: Supports replicated or sharded parameter delivery and gradient reduction, with configurable route choices. - **Pipeline Overlap**: Overlaps data movement and CPU updates with GPU computation to maximize efficiency. ### Tech stack: Python ### Why it matters: SlideDP enables the training of large language models (LLMs) on multiple GPUs, addressing the limitations of GPU memory by optimizing parameter delivery, gradient aggregation, and CPU updates. This approach allows for efficient fine-tuning of models beyond the constraints of individual GPUs, making it particularly valuable for scaling training across multiple GPUs.

Overview: SnowLLM is a high-performance local inference framework optimized for AMD Ryzen AI Max and compatible AMD RDNA 3/3.5 APUs. It features hand-tuned kernels, low-concurrency inference, and adaptive speculative decoding for improved performance. Key features: - Hand-tuned kernels for Ryzen AI Max series - Precision optimized for hardware performance - Low-concurrency inference for better performance - Speculative decoding for faster responses Tech stack: Python Why it matters: SnowLLM offers significant performance improvements over llama.cpp, achieving up to 9x faster prefill and 2x faster decode, making it ideal for applications requiring high inference speed and low resource usage. Repository: SnowLLM/SnowLLM GitHub stars: 19 GitHub description: Your Ryzen AI Max is running at half speed. SnowLLM: up to 9x faster prefill and 2x faster decode than llama.cpp on the same GGUF. Qwen3.6-35B-A3B, Qwen3.8-Flash-Next, DeepSeek-V4-Flash.

Overview: ABC is an open-source multimodal embedding model designed to give users maximum control over how images are represented in embeddings. It achieves outstanding performance in tasks like VQA and zero-shot classification. Key features: - 📝 Maximum control over image representation in embeddings - 📓 Pretrained on a large dataset of difficult embedding samples - 📊 Outputs high-quality embeddings for various tasks Tech stack: Python Why it matters: ABC is a powerful tool for controlling image representations in embeddings, making it ideal for applications requiring precise control over image features.

### Overview: VISTA is a video spatiotemporal augmentation method presented in the CVPR 2025 paper "VISTA: Enhancing Long-Duration and High-Resolution Video Understanding by Video Spatiotemporal Augmentation". It generates long-duration and high-resolution video instruction-following data to improve the video understanding capabilities of video language models (LMMs). ### Key Features: - **Video Spatiotemporal Augmentation**: VISTA combines multiple videos spatially and temporally to create augmented video samples with longer durations and higher resolutions. - **Instruction Following Data**: The method synthesizes instruction data based on these augmented videos, aiming to enhance the long and high-resolution video understanding capabilities of video LMMs. - **High-Quality Dataset**: VISTA-400K, a high-quality video instruction-following dataset, is created using existing public video-caption datasets, making it fully open-sourced and scalable. ### Tech Stack: Python ### Why it Matters: VISTA provides a novel approach to improving the understanding of long-duration and high-resolution videos by augmenting them with synthetic data. This method can significantly enhance the performance of video language models, making them better suited for tasks that require detailed video understanding.