Projects

Overview: MenteDB is a ground-up Rust database engine designed specifically for AI agent memory, built with advanced features like WAL, HNSW, knowledge graphs, and speculative context pre-assembly. Key features: - Advanced indexing and query capabilities - Support for knowledge graphs - Speculative context pre-assembly for improved performance - Built-in support for AI/LLM consumption patterns Tech stack: Rust Why it matters: MenteDB is a purpose-built database engine that understands the unique needs of AI/LLM agents, offering advanced indexing, knowledge graph support, and speculative context pre-assembly to optimize performance and memory usage. Repository: nambok/mentedb GitHub stars: 122 GitHub description: A cognition aware database engine for AI agent memory. Not a wrapper, a ground up storage engine that thinks.

Overview: This library simplifies the process of running few-shot prompting experiments on large-language models by providing a straightforward interface and additional features like rate-limiting and retry logic. Key features: - A library for hassle-free few-shot prompting. - Support for various tasks such as boolean expression simplification, code completion, and more. - Includes rate-limiting and retry logic for better performance and reliability. - Easy to use with Python and can be run from the command line. Tech stack: Python Why it matters: This library streamlines the process of experimenting with few-shot prompting, making it accessible and efficient for developers and researchers working with large-language models.

Overview: This repository contains the code for the NAACL 2022 paper "Good Visual Guidance Makes A Better Extractor: Hierarchical Visual Prefix for Multimodal Entity and Relation Extraction". The model architecture is described in the provided image, and it uses hierarchical modal fusion to improve multimodal entity and relation extraction. Key features: - Hierarchical Visual Prefix for Multimodal Entity and Relation Extraction - Code for the NAACL 2022 paper - Model architecture illustrated - Data preprocessing steps for visual object detection and image-to-object mapping Tech stack: Python Why it matters: This repository presents a novel approach to multimodal entity and relation extraction using hierarchical visual prefix, which could enhance the accuracy and efficiency of extracting information from text and images. GitHub stars: 124

### Overview: This repository contains the source code and documentation for the paper "Dispersion loss counteracts embedding condensation and improves generalization in small language models" presented at ICML 2026. The paper observes and addresses the issue of embedding condensation in small language models, proposing dispersion loss as a solution to improve model generalization. ### Key Features: - **Dispersion Loss**: A training objective designed to counteract embedding condensation in small language models. - **Observation of Embedding Condensation**: The paper identifies that smaller language models exhibit more severe embedding condensation, with token embeddings collapsing into a narrow subspace. - **Robustness to Dataset Choice**: The effect of embedding condensation is found to be robust to changes in input datasets, indicating a universal phenomenon. ### Tech Stack: - **Programming Languages**: Python ### Why It Matters: The proposed dispersion loss objective addresses a significant issue in the training of small language models, improving their generalization and performance. The paper's findings and the accompanying code provide a valuable contribution to the field of natural language processing and language model training.

Overview: This repository offers a unified interface for guiding various generative models, including 2D/3D GANs, diffusion models, and their variants, using CLIP, classifiers, and face IDs. The unified approach leverages latent-space energy-based models to condition generative models, enabling a plug-and-play manner of guiding them. Key features: - Unified interface for guiding generative models - Integration with CLIP, classifiers, and face IDs - Plug-and-play guidance for various generative models Tech stack: Python Why it matters: This unified approach simplifies the process of guiding generative models, making it easier to condition them with different types of inputs, thus enhancing their applications in various domains. Repository: ChenWu98/unified-generative-zoo GitHub stars: 123 GitHub description: [ICCV 2023] https://arxiv.org/abs/2210.05559

Overview: S-Eval is a comprehensive and automated safety evaluation tool designed specifically for large language models. It provides a standardized benchmark and evaluation metrics to assess the safety of these models. Key features: - Comprehensive evaluation of large language models - Supports multiple language versions - Offers various benchmark scales for different evaluation needs - Provides detailed risk prompts and attack scenarios Tech stack: Python, Hugging Face, and other open-source tools Why it matters: S-Eval helps researchers and developers evaluate the safety of large language models accurately and consistently, ensuring the responsible and ethical use of these models in various applications.

### Overview: SLED is a novel factuality decoding approach designed for large language models. It leverages the latent knowledge within LLMs by contrasting the final layer's logits with early layers' logits, enabling the self-evolution of the output distribution to align with real-world facts. ### Key Features: - **Model Versatility**: Compatible with most large language model families. - **Task Versatility**: Tested with factual accuracy enhancement across various tasks and benchmarks. - **High Compatibility**: Can be flexibly combined with other decoding methods. - **High-Quality Outputs**: Reduces repetition and ensures fluent responses. - **Negligible Computational Overhead**: Minimal additional costs, suitable for real-time use. - **Interpretability**: Provides insights into inference-time computing algorithms. ### Tech Stack: Python ### Why It Matters: - **Model Versatility**: Compatible with most large language model families. - **Task Versatility**: Tested with factual accuracy enhancement across various tasks and benchmarks. - **High Compatibility**: Can be flexibly combined with other decoding methods. - **High-Quality Outputs**: Reduces repetition and ensures fluent responses. - **Negligible Computational Overhead**: Minimal additional costs, suitable for real-time use. - **Interpretability**: Provides insights into inference-time computing algorithms. ### Repository: JayZhang42/SLED ### GitHub Stars: 123 ### GitHub Description: SLED: Self Logits Evolution Decoding for Improving Factuality in Large Language Models [NeurIPS 2024]

Overview: ElevateAI Java SDK provides a comprehensive API for speech-to-text, behavioral analysis, and sentiment analysis of voice interactions, making it easier for developers to integrate AI into their applications. Key features: - Speech-to-text (ASR) functionality - Behavioral analysis - Sentiment analysis of voice interactions Tech stack: Java Why it matters: ElevateAI's Java SDK simplifies the integration of AI capabilities into Java-based projects, enabling developers to leverage advanced speech and interaction analysis without needing deep expertise in AI technologies. Repository: NICEElevateAI/ElevateAIJavaSDK Primary language: Java GitHub stars: 123 GitHub description: Java SDK for ElevateAI

Overview: Sounder is an intent recognizing algorithm designed to predict the intent of a given text. It is an abstraction of the Sounder Algorithm, which is explained in detail in the provided paper. Key features: - Instantiation of the class with a dataset - Search method for finding the most probable index in the dataset - Probability method for calculating the likelihood of a query matching the dataset - Filtering method for refining the search results Tech stack: Python Why it matters: Sounder is useful for applications that require intent recognition, such as chatbots, customer support systems, or any scenario where understanding the user's intent is crucial.

### Overview: RealCompo is a training-free and transferred-friendly text-to-image generation framework that enhances both realism and compositionality in generated images. ### Key Features: - **RealCompo** is a new training-free and transferred-friendly text-to-image generation framework. - It leverages the strengths of text-to-image models and spatial-aware image diffusion models to improve both realism and compositionality in generated images. - An intuitive balancer dynamically balances the strengths of the two models during the denoising process, allowing for plug-and-play use of any model. ### Tech Stack: Python ### Why It Matters: RealCompo consistently outperforms state-of-the-art text-to-image models and spatial-aware image diffusion models in multiple-object compositional generation while maintaining satisfactory realism and compositionality of the generated images.

### Overview: This project combines the features of FastChat and Alpaca-Lora to create a low-end GPU capable model for fine-tuning LLaMA, using the ShareGPT dataset. The repository includes everything needed for training, inference, and exporting the model. ### Key Features: - Support for multilingual and multi-round chat sessions. - Training and inference capabilities on low-end GPUs using LORA. - Open-source dataset, training code, and export model code. - Latest model (Sunlight 13B) available for download. ### Tech Stack: Jupyter Notebook ### Why it Matters: This project aims to produce a model similar to FastChat but with significantly cheaper hardware requirements, particularly for non-Ampere GPUs.

Overview: This project provides a Keras interface for Kaldi ASR, simplifying the integration of Kaldi tools with Keras for speech recognition tasks. Key features: - Trains Deep Learning models using Kaldi's GMM system - Works with standard Kaldi data and alignment directories - Supports mini-batch training - Supports LSTM, maxout, and dropout training - Easily extendable to other deep learning implementations in Keras - Decodes test utterances in Kaldi style Tech stack: Python Why it matters: This code simplifies the integration of Kaldi tools with Keras, making it easier to build and test deep learning models for speech recognition tasks. Repository: dspavankumar/keras-kaldi GitHub stars: 123

### Overview: FLowHigh: Towards Efficient and High-Quality Audio Super-Resolution with Single-Step Flow Matching is a project aimed at improving audio super-resolution techniques. The project includes an efficient audio super-resolution model that can reconstruct low-resolution audio into high-resolution audio at 48 kHz. The model is designed to be both efficient and high-quality, making it suitable for various applications in audio processing and enhancement. ### Key Features: - **Efficient Model Design**: The FLowHigh model is designed to be highly efficient and high-quality, making it suitable for real-time applications. - **Single-Step Flow Matching**: The model uses a single-step flow matching approach to achieve high-quality results, which is particularly useful in scenarios where computational resources are limited. - **Audio Super-Resolution**: The project focuses on improving audio super-resolution techniques, enabling the reconstruction of low-resolution audio to high-resolution audio at 48 kHz. ### Tech Stack: The project utilizes Python as the primary programming language, leveraging libraries such as PyTorch for deep learning and TensorFlow for data preprocessing and model training. ### Why It Matters: The FLowHigh project addresses the need for efficient and high-quality audio super-resolution techniques, particularly in scenarios where computational resources are limited. By focusing on a single-step flow matching approach, the model significantly reduces computational overhead, making it suitable for real-time applications and edge devices. The project also contributes to the field of audio processing by providing a robust and efficient solution for audio super-resolution tasks. ### Repository: The repository for this project is located at: [jjunak-yun

### Overview: This repository documents the journey of the community in achieving the goal of making Large Language Models (LLMs) produce reliable reasoning and make decisions that are reason-responsive. ### Key features: - **Deliberative Prompting**: Techniques to elicit reasoning traces from LLMs. - **Chain-of-Thought (CoT)**: Method to improve the quality of reasoning output. - **Self-Correction**: Mechanism to correct errors in reasoning. - **Reasoning Analytics**: Tools to analyze and improve reasoning capabilities. - **Datasets**: Resources for training and testing reasoning models. - **Tools and Frameworks**: Implementations of various reasoning techniques. - **Other Resources**: Additional reading materials and community insights. ### Tech stack: Python, Jupyter Notebook, and various NLP libraries. ### Why it matters: Deliberative prompting and reasoning techniques have become essential for improving the reliability and reason-responsiveness of LLMs, making them more effective in complex tasks and decision-making processes.

Overview: This project is an end-to-end speech recognition system that leverages the Connectionist Temporal Classification (CTC) cost function, inspired by Baidu's Deep Speech papers. The system is trained on a combined corpus containing over 900 hours of audio data and achieves an error rate of 12.6% on the test dataset. Key features: - End-to-end speech recognition system - Uses CTC cost function - Trained on a combined corpus of 900+ hours of audio data - Achieves a word error rate (WER) of 12.6% on the test dataset Tech stack: Python Why it matters: This system is designed to be highly efficient and effective for speech recognition tasks, making it suitable for various applications requiring real-time or near-real-time speech processing. Repository: mdangschat/ctc-asr GitHub stars: 123 GitHub description: End-to-end trained speech recognition system, based on RNNs and the connectionist temporal classification (CTC) cost function.

### Overview: This repository serves as a collection of resources for practitioners to explore and fine-tune Large Language Models (LLMs). It includes tools, software, libraries, datasets, research papers, and more, designed to help individuals and organizations get started with LLMs. ### Key Features: - Leaderboards for tracking and evaluating LLMs and chatbots. - Kaggle & Colab Notebooks for practical examples and tutorials. - Courses and lectures on LLM concepts and techniques. - Blogs and articles for in-depth discussions and insights. ### Tech Stack: Python and Jupyter Notebooks. ### Why it Matters: This repository provides a comprehensive collection of resources for anyone interested in diving into the world of LLMs. Whether you're a researcher, developer, or simply curious about the latest advancements in AI, this list offers valuable information and practical examples to help you get started.

### Overview: RSTNet is an advanced model designed for captioning tasks that incorporates adaptive attention mechanisms for both visual and non-visual words. This project provides the official code for the paper "RSTNet: Captioning with Adaptive Attention on Visual and Non-Visual Words" presented at CVPR 2021. ### Key features: - **Captioning with Adaptive Attention**: The model uses adaptive attention to improve the accuracy of caption generation by considering both visual and non-visual information. - **Visual and Non-Visual Words**: The model effectively handles captions that include both visual and non-visual elements, enhancing its ability to understand complex scenes. - **Adaptive Attention Mechanism**: The model employs an adaptive attention mechanism to focus on relevant parts of the input data, improving the quality of the generated captions. ### Tech stack: The project is built using Python, leveraging the libraries `torch`, `transformers`, and `spacy` for natural language processing and image processing tasks. ### Why it matters: RSTNet demonstrates significant improvements in captioning tasks by effectively integrating visual and non-visual information, making it a valuable tool for researchers and practitioners in the field of computer vision and natural language processing. ### Repository: zhangxuying1004/RSTNet - **Stars**: 123 - **Description**: Official Code for 'RSTNet: Captioning with Adaptive Attention on Visual and Non-Visual Words' (CVPR 2021) - **Primary Language**: Python - **GitHub Description**: "Official Code for

Overview: PyRIT is an open-source framework designed to help security professionals and engineers identify risks in generative AI systems. Key features: - Risk identification - Proactive approach - Framework for generative AI Tech stack: Python Why it matters: PyRIT provides a comprehensive tool for detecting potential risks in AI systems, ensuring robust security practices. Repository: Azure/PyRIT Primary language: Python GitHub stars: 123 GitHub description: The Python Risk Identification Tool for generative AI (PyRIT) is an open source framework built to empower security professionals and engineers to proactively identify risks in generative AI systems.

### Overview: 《深度学习：基础与概念》是一本全面的教材，涵盖了深度学习的理论基础和实际应用。本书由Christopher M. Bishop和Hugh Bishop合著，由Springer于2024年出版。该书提供从基础概念到高级架构的全面覆盖，包括概率论、现代架构、注意力机制、GAN、VAE等主题。本书适用于机器学习初学者、学术研究人员、自学者和专业人士。 ### Key features: - **书籍内容全面覆盖**：从基础概念到高级架构，全面涵盖深度学习理论和应用。 - **概率论介绍**：详细讲解概率论，为深度学习提供坚实的数学基础。 - **现代架构深入探讨**：深入探讨MLP、CNN、RNN、Transformer、GNN等现代架构。 - **习题解答**：提供概率论、标准分布、单层网络、Transformer、变分自编码器等主题的数学推导解答。 - **数据集和练习**：包含用于手写数字识别的MNIST数据集和涵盖手写数字识别、PyTorch入门、神经网络实现（包括UNet架构）等主题的Jupyter笔记本。 ### Tech stack: - **Jupyter Notebook**：用于编写和展示深度学习相关的代码和练习。 ### Why it matters: 本书是深度学习领域的重要资源，提供了全面的理论基础和实际应用案例，适合各种学习和研究需求。通过本书，读者可以系统地学习深度学习的基础知识和高级应用，为未来的研究

### Overview: The `cavegemma` project fine-tunes Gemma 4 31B to speak caveman-mode natively, reducing output tokens by 27% across 193 holdout pairs. The model preserves cosine similarity at 0.91-0.98 and reproduces code fences byte-exact 96-100% of the time. ### Key features: - **27% fewer output tokens across 193 holdout pairs.** - **Code fences reproduced byte-exact 96-100% of the time.** - **Fidelity held while all that came off.** ### Tech stack: Python ### Why it matters: This project showcases the effectiveness of fine-tuning models to achieve specific conversational styles, demonstrating how to train a model to speak in a caveman mode without compromising fidelity or code preservation.