Project directory

Projects

Open-source AI projects tracked by the leaderboard, ordered by GitHub stars.

Jupyter Notebook 38 stars

mddunlap924/StableDiffusion2-Image-to-Text

Stable Diffusion with Text-to-Image and Image-to-Text

Overview: This repository contains Python code for fine-tuning models to predict prompts and/or embeddings from generated images, focusing on the reverse process of predicting text prompts from images. Key features: - Predicting text prompts from images using CoCa model - Predicting text embeddings from images using OpenAI CLIP-ViT model Tech stack: Python, Jupyter Notebook Why it matters: This work is part of the Kaggle Stable Diffusion - Image to Prompts competition, where the goal is to predict the text prompts used to generate images. The repository provides a framework for fine-tuning various models to predict the text prompt from generated images, making it useful for tasks such as prompt engineering in the context of image generation models like Stable Diffusion.

Jupyter Notebook 38 stars

naxingyu/interactive_e2e_speech_recognition

Overview: This repository is designed as a beginner-friendly tutorial on end-to-end speech recognition, focusing on a simple and interactive approach using YesNo data. Key features: - Interactive CTC trainer - YesNo data for training Tech stack: Jupyter Notebook Why it matters: This project aims to provide a straightforward and accessible introduction to end-to-end speech recognition, making it ideal for beginners with basic knowledge of the topic. Repository: naxingyu/interactive_e2e_speech_recognition Primary language: Jupyter Notebook GitHub stars: 38

Python 38 stars

neuralwork/instruct-finetune-mistral

Fine-tune Mistral 7B to generate fashion style suggestions

### Overview: Fine-tuning Mistral 7B to generate fashion style suggestions. ### Key features: - Fine-tuning Mistral 7B model for fashion style suggestions. - Semi-manually crafted dataset for training. - Parameter-efficient fine-tuning (PEFT) using LoRA and 4-bit quantization. - Fine-tuned model saved for future use. ### Tech stack: Python ### Why it matters: This project provides a practical guide on fine-tuning large language models like Mistral 7B, specifically for generating fashion style suggestions. It demonstrates how to efficiently train and fine-tune models using parameter-efficient techniques, making it useful for developers looking to optimize their models for specific tasks.

Python 38 stars

paulocoutinhox/mini-llm

Simple and lightweight tool to fine-tune GPT models (like GPT-2 and GPT-Neo) using your own data — built with Python and Transformers. Adapt powerful language models to your domain with ease.

### Overview: Mini LLM (SLM) is a simple and lightweight tool designed to fine-tune pre-trained language models like GPT-2 and GPT-Neo using your own data. This project allows you to adapt powerful language models to your specific domain or language while leveraging their existing knowledge. ### Key Features: - **Simple Command-Line Interface**: Provides a straightforward way to generate text based on prompts. - **Fine-Tuning Capabilities**: Supports custom data for fine-tuning pre-trained models. - **Multiple Pre-Trained Models**: Compatible with GPT-Neo and GPT-2, among others. - **Configurable Parameters**: Allows for easy modification and customization. - **Python and Transformers Library**: Built with Python and the Transformers library for efficient processing. - **Support for Python 3.9 or Later**: Ensures compatibility with the latest Python versions. - **Comparison of Original and Fine-Tuned Models**: Helps in understanding the influence of your training data on the model's output. ### Tech Stack: - Python - Transformers Library ### Why It Matters: This tool simplifies the process of adapting language models to your specific needs, making it easier to leverage the models' existing knowledge while tailoring them to your domain or language. It's particularly useful for developers looking to enhance their models without the need for extensive machine learning knowledge. ### Repository: paulocoutinhox/mini-llm - GitHub Stars: 38 - GitHub Description: A simple and lightweight tool to fine-tune GPT models (like GPT-2 and G

Jupyter Notebook 38 stars

redis-developer/gcp-redis-llm-stack

Reference architecture for LLM-based applications on Google Cloud Platform with Redis Enterprise as a high-performance data layer.

### Overview: This repository provides a foundational architecture for building Large Language Model (LLM) applications using Redis Enterprise and Google Cloud Platform (GCP) services. It leverages Vertex AI for deploying and managing LLM instances and Google's powerful language models. ### Key features: - **Semantic Caching**: Improves LLM performance and associated costs by caching semantic data. - **Vector Search**: Facilitates fast and scalable vector search, enhancing context retrieval from a knowledge base. - **Large Language Models**: Utilizes `text-embedding-gecko@003` for embeddings and `gemini-1.5-flash-001` for LLM generation and chat. ### Tech stack: Jupyter Notebook ### Why it matters: This reference architecture serves as a practical guide for developers looking to build scalable and high-performance LLM applications using Redis and GCP services.

Python 38 stars

romitjain/kachua-mlsys

[MLSys 26] 🥇 Solution for Gated Delta Net Track of MLSys 26 Flash infer competition

Overview: This repository contains optimized Gated Delta Network (GDN) kernels for the MLSys 2026 FlashInfer AI Kernel Generation Contest, targeting NVIDIA B200. It includes two active definitions: `gdn_prefill_qk4_v8_d128_k_last` and `gdn_decode_qk4_v8_d128_k_last`, each with their own scripts and configurations. Key features: - Prefill kernel for processing multiple prompt tokens using chunked recurrence blocks. - Decode kernel for handling one token at a time with low-latency single-token execution. Tech stack: Python. Why it matters: This repository provides optimized kernels for the MLSys 2026 FlashInfer AI Kernel Generation Contest, targeting NVIDIA B200, which is crucial for improving the performance of AI models in real-time applications. Repository: romitjain/kachua-mlsys, with 37 GitHub stars, showcasing the importance of efficient AI kernel optimization in competitive AI competitions.

Python 38 stars

ronb1964/TuxTalk

Free push-to-talk voice dictation for Linux. Hold a key, talk, let go. Works on Wayland and X11, runs offline with Parakeet or Whisper, and fixes "no wait" slips with an optional local AI.

### Overview: TuxTalk is a free voice dictation tool for Linux that allows you to dictate text by holding a key and talking. It works offline and supports Wayland and X11, including GNOME, KDE, COSMIC, Sway, and X11. The app fixes "no wait" slips with optional local AI. ### Key Features: - **Dictation Mode:** Hold a key, talk, let go. - **Offline Functionality:** Runs completely offline. - **Voice Detection:** A small voice detector built into the app. - **Noise Reduction:** Handles various types of background noise effectively. - **Automatic Send:** Sends dictation text with a click or press of the "send" button. - **Customization:** Supports various keyboard layouts and hotkey setups. ### Tech Stack: Python ### Why It Matters: TuxTalk is a practical and user-friendly tool for dictating text on Linux, making it an essential addition to any developer's toolkit. Its ability to work offline and its effective noise reduction make it a standout choice for those who need to dictate text frequently. ### Repository: Repository: ronb1964/TuxTalk GitHub Stars: 36 GitHub Description: Free voice dictation for Linux. Hold a key, talk, let go. Works on Wayland and X11, runs offline with Parakeet or Whisper, and fixes "no wait" slips with an optional local AI. ### Installation: - Download from the GitHub releases page. - Install using the AUR package manager. ### What It Does

Jupyter Notebook 38 stars

sail-sg/D-TRAK

Intriguing Properties of Data Attribution on Diffusion Models (ICLR 2024)

### Overview: This project explores the intriguing properties of data attribution in diffusion models, presenting counter-intuitive observations that outperform previous baselines by a significant margin. The findings challenge the conventional wisdom on design choices for attributing diffusion models, offering valuable insights into their performance. ### Key Features: - **Data Attribution on Diffusion Models:** The project focuses on attributing diffusion models, providing counter-intuitive observations that outperform previous baselines. - **Intriguing Properties:** The findings reveal counterfactuals that are not aligned with theoretical expectations, leading to improved performance. - **GitHub Repository:** The project is hosted on GitHub at `https://github.com/sail-sg/D-TRAK`, with 38 GitHub stars. ### Tech Stack: - **Primary Language:** Jupyter Notebook ### Why It Matters: The project provides novel insights into the performance of diffusion models, challenging the conventional wisdom on attribution methods and offering a significant improvement over existing baselines.

Python 38 stars

space-bacon/SRT

Adapters and probes that read frozen LLM hidden states, and measurements of how that read-out carries across models, precision, hardware and modality.

### Overview: SRT — Semiotic-Reflexive Transformer (Adapter Architecture) is a lightweight module that injects semiotic awareness into any frozen causal language model. It reads divergence from hidden states, tracks reflexive awareness, and optionally injects semiotic corrections back into the stream. The structure it reads is invariant across host scale, weight precision, and hardware, making it portable and lightweight. ### Key Features: - **Reads Divergence from Hidden States:** SRT taps observe divergence from hidden states, allowing for the extraction of semiotic signals. - **Tracks Reflexive Awareness:** SRT tracks reflexive awareness, providing insights into the model's internal mechanisms. - **Injects Semiotic Corrections:** SRT optionally injects semiotic corrections back into the stream, enhancing the model's understanding and interpretation. ### Tech Stack: - **Python** ### Why It Matters: SRT is a powerful tool for instrumenting frozen language models, making them more semiotic-aware and adaptable across different deployment tiers and hardware. It allows for lightweight, portable instrumentation that trains in hours with minimal impact on the base model, making it ideal for research and development environments. ### Repository: The SRT repository is hosted on GitHub at [space-bacon/SRT](https://github.com/space-bacon/SRT). ### GitHub Stars: The repository has been starred by 38 developers, indicating its growing interest and utility in the community. ### The Browser Tier: A small model reads a large one, with a 382 MB Qwen3-0.6B backbone and a

Python 38 stars

taraak/pita

Code for the paper Progressive Inference-Time Annealing of Diffusion Models for Sampling from Boltzmann Densities.

Overview: This repository contains the code for the paper "Progressive Inference-Time Annealing of Diffusion Models for Sampling from Boltzmann Densities," which proposes a method to iteratively sample from a series of temperature annealed Boltzmann distributions using diffusion models. Key features: - Code for the paper "Progressive Inference-Time Annealing of Diffusion Models for Sampling from Boltzmann Densities" - Repository built on the repository of "Iterated Denoising Energy Matching for Sampling from Boltzmann Densities" - Experimented on Lennard-Jones system of 13 particles and small peptide molecules (Alanine Dipeptide and Tripeptide) Tech stack: Python Why it matters: This research aims to improve the sampling process of Boltzmann distributions, which is crucial in various applications such as molecular dynamics simulations and machine learning models.

Python 38 stars

thuml/dynamical-diffusion

About Code release for "Dynamical Diffusion: Learning Temporal Dynamics with Diffusion Models" (ICLR 2025)

### Overview: Dynamical Diffusion is a novel approach to learning temporal dynamics using diffusion models, presented at ICLR 2025. ### Key Features: - **Training**: A script to train a model (e.g., DyDiff for Turbulence) using the provided configuration file. - **Sampling**: A script to generate samples from a trained model. - **Evaluation**: A script to evaluate the generated samples. ### Tech Stack: Python ### Why It Matters: Dynamical Diffusion offers a unique method for learning temporal dynamics, which is particularly useful in applications requiring temporal information, such as turbulence modeling.

Unknown 38 stars

unikcc/AwesomeEventExtraction

Event extraction survey and paper list

### Overview: This repository is a curated survey and paper list for event extraction across various domains, including text, speech, vision, and multimodal streams. It provides a comprehensive overview of the field, categorizing subtasks and methodological advancements from traditional rule-based systems to modern Large Language Models and Multimodal LLMs. ### Key features: - Comprehensive survey of event extraction papers across text, speech, vision, and multimodal streams. - Official code and benchmarks for evaluation. - Detailed categorization of subtasks and methodological advancements. - Updated automatically with the latest papers and metrics. ### Tech stack: GitHub stars: 38 ### Why it matters: This repository serves as a valuable resource for researchers, developers, and practitioners interested in event extraction, providing a clear roadmap of the field's evolution and current state.

Python 38 stars

upskyy/ContextNet

PyTorch implementation of "ContextNet: Improving Convolutional Neural Networks for Automatic Speech Recognition with Global Context" (INTERSPEECH 2020)

### Overview: ContextNet is a PyTorch implementation of a context-aware convolutional neural network (ContextNet) designed for automatic speech recognition. It features a fully convolutional encoder that incorporates global context information into convolution layers using squeeze-and-excitation modules. The model supports three size models: small, medium, and large, each with a different number of channels controlled by the global parameter alpha. This repository contains only the model code and provides instructions for training and using the ContextNet model. ### Key Features: - **Model Architecture:** - **Configuration of the Encoder:** The encoder's number of channels is determined by the global parameter alpha, which can be adjusted based on the model size (small, medium, large). - **Convolution Block Architecture:** Each convolution block has five layers except the first and last, which have a stride of one and two, respectively. - **1D Squeeze-and-Excitation (SE) Module:** The SE module reduces the convolution result into a 1D vector and applies a Sigmoid function to map it to (0, 1), followed by pointwise multiplications to enhance feature extraction. ### Tech Stack: - **Programming Languages:** Python ### Why It Matters: ContextNet is designed to improve the performance of convolutional neural networks (CNNs) in automatic speech recognition by incorporating global context information. This makes it particularly useful for tasks where context is crucial, such as speech recognition and language modeling. ### Repository: - **Repository:** upskyy/ContextNet - **GitHub Stars:** 38 -

Python 38 stars

vortezwohl/Bhakti

An easy-to-use vector database.

### Overview: Bhakti is a lightweight vector database designed to meet the storage and semantic search needs of small and medium-sized datasets. It supports various similarity calculation methods and a domain-specific language (DSL) for document-based pattern matching pre-filtering, making it easy to migrate data and integrate seamlessly with Python3. ### Key features: - Supports a variety of similarity calculation methods. - Domain-specific language (DSL) for document-based pattern matching pre-filtering. - Portable data files for easy migration. - Flexible data management and seamless integration with Python3. - Memory-enhanced large language model dialogue solution for fine-grained control over semantic importance. ### Tech stack: Python ### Why it matters: Bhakti is designed to empower large language models with semantic search capabilities, making it a valuable tool for applications requiring efficient and accurate data processing and retrieval. Its lightweight nature makes it particularly suitable for scenarios involving small and medium-sized datasets, offering a clear advantage in terms of processing speed and resource efficiency.

Python 38 stars

xmed-lab/FDDM

[MICCAI 2023] Fundus-Enhanced Disease-Aware Distillation Model for Retinal Disease Classification from OCT Images

Overview: This project is a PyTorch implementation of a paper titled "Fundus-Enhanced Disease-Aware Distillation Model for Retinal Disease Classification from OCT Images" presented at the MICCAI 2023 conference. It focuses on classifying retinal diseases using OCT images and fundus images. Key features: - A new dataset, TOPCON-MM, is used for training, validation, and testing. - The project includes data preprocessing steps such as contrast-limited adaptive histogram equalization for fundus images and 3x3 median filtering for OCT images. - The implementation includes a structure for organizing the data folder, including train, val, and test collections. - The project also includes a script to train the model using pre-trained fundus single-modal model and a fundus-enhanced OCT model with specific parameters. Tech stack: Python, PyTorch, Scikit-learn, NumPy, OpenCV, Pillow, Matplotlib, and Hugging Face Transformers. Why it matters: This project is significant for researchers and developers interested in medical imaging and machine learning, as it demonstrates a practical application of distillation methods in retinal disease classification from OCT images.

Python 38 stars

youngharold/tightwad

Mixed-vendor GPU inference cluster manager with speculative decoding

Overview: Tightwad is a mixed-vendor GPU inference cluster manager that accelerates inference by leveraging speculative decoding across network-separated servers. This allows for faster model processing with reduced latency. Key features: - Mixed-vendor GPU inference cluster management - Speculative decoding proxy - Pools CUDA and ROCm GPUs across machines - Accelerates inference via application-layer speculative decoding Tech stack: Python Why it matters: Tightwad enables faster model processing by leveraging speculative decoding across multiple GPUs, making it suitable for applications requiring high inference speed and reduced latency. Repository: youngharold/tightwad Primary language: Python GitHub stars: 38 GitHub description: Mixed-vendor GPU inference cluster manager with speculative decoding README excerpt: # Tightwad [![PyPI](https://img.shields.io/pypi/v/tightwad)](https://pypi.org/project/tightwad/) [![CI](https://github.com/youngharold/tightwad/actions/workflows/ci.yml/badge.svg)](https://github.com/youngharold/tightwad/actions) [![License](https://img.shields.io/github/license/youngharold/tightwad)](LICENSE) [![Python](https://img.shields.io/pypi/pyversions/tightwad)](https://pypi.org/project/tightwad/) Mixed-vendor GPU inference cluster manager with speculative decoding proxy. Pools CUDA and ROCm GPUs across machines using [llama.cpp RPC](https://

Python 38 stars

zengxianyu/structured-noise

Structured Noise Generation from the paper NeuralRemaster with Phase-Preserving Diffusion

Overview: Structured Noise Generation is a project that leverages the NeuralRemaster paper to generate structured noise using Phase-Preserving Diffusion, focusing on applications in ComfyUI and Command Line Interface (CLI) environments. Key features: - Phase-Preserving Diffusion-based structured noise generation - Compatible with ComfyUI and CLI environments - Supports training and inference Tech stack: Python Why it matters: This project offers a practical application of advanced diffusion models for generating structured noise, making it useful in various creative and technical domains. Repository: zengxianyu/structured-noise Primary language: Python GitHub stars: 38 GitHub description: Structured Noise Generation from the paper NeuralRemaster with Phase-Preserving Diffusion

TypeScript 38 stars

zhebrak/llm-cluster-simulator

Simulate distributed LLM training and inference across GPU clusters in your browser. Memory, throughput, cost, and parallelism strategy planning.

### Overview: A comprehensive tool for simulating distributed LLM training and inference across GPU clusters, providing accurate estimates of memory, throughput, cost, and parallelism strategies without the need for actual GPU provision. ### Key features: - Simulates the performance of any model-parallelism combination across different GPU configurations. - Estimates MFU, memory, throughput, and cost for various model-parallelism setups. - Automatically optimizes parallelism layouts for optimal performance. - Supports training and inference scenarios with features like speculative decoding, continuous batching, and mixed precision. ### Tech stack: TypeScript ### Why it matters: This simulator replaces traditional methods of estimating cluster sizes with a physics-based model, offering a more accurate and efficient way to plan distributed training and inference setups. It is particularly useful for evaluating strategies, validating cluster budgets, and building intuition about parallelism trade-offs, especially for those not familiar with profiling production workloads.

Python 37 stars

AlonzoLeeeooo/LCDG

The official code implementation of "LaCon: Late-Constraint Diffusion for Steerable Guided Image Synthesis".

Overview: LaCon: Late-Constraint Diffusion for Steerable Guided Image Synthesis is an official code implementation of a paper that proposes a novel approach called Late-Constraint Diffusion (LaCon) for steerable guided image synthesis. This method integrates various conditions into pre-trained diffusion models and aligns them with the internal features of the models to guide the sampling process, resulting in more controllable and generalizable image synthesis. Key features: - LaCon integrates various conditions into pre-trained diffusion models. - It establishes an alignment between external conditions and the internal features of diffusion models to guide sampling. - The pre-trained model weights are available on Huggingface and ModelScope. Tech stack: Python Why it matters: LaCon offers a more flexible and efficient solution for steerable guided image synthesis by providing controllable and generalizable image generation through the integration of various conditions into pre-trained models. This approach addresses the limitations of early-constraint methods by enabling more flexible and efficient solutions for diffusion models. Repository: AlonzoLeeeooo/LCDG Primary language: Python GitHub stars: 37 GitHub description: The official code implementation of "LaCon: Late-Constraint Diffusion for Steerable Guided Image Synthesis". README excerpt:  # Table of Contents - [1. News](#news) - [2. To-Do Lists](#to-do-lists) - [3. Overview of LaCon](#overview-of-lacon) -

TypeScript 37 stars

AmirLayegh/procurement-ai-assistant

🏆 AI Procurement Assistant - Google Cloud Sweden Hackathon Winner. Natural language search for retail buyers using Superlinked semantic search.

Overview: Procurement AI Assistant is an AI-powered semantic search tool designed to help retail buyers make informed purchasing decisions quickly and efficiently. Key features: - 🗣️ Natural Language Search - Queries using plain English. - ⚡ Instant Results - Sub-3-second response times. - 📊 Smart Analytics - Supplier reliability, profit margins, return rates. - 🎯 Dual Purpose - Serves both procurement and marketing teams. - 🔍 Vector Search - Powered by advanced semantic understanding. Tech stack: TypeScript, Python 3.11, Superlinked Framework, FastAPI, Qdrant Vector Database, Google Cloud Run, Firebase Hosting, Google Cloud Platform, OpenAI GPT-4, Sentence Transformers, Vector Search. Why it matters: This tool empowers retail buyers to make intelligent procurement decisions by providing instant, actionable insights through natural language search, streamlining the decision-making process and reducing the need for lengthy data requests. Repository: AmirLayegh/procurement-ai-assistant GitHub stars: 37 GitHub description: 🏆 AI Procurement Assistant - Google Cloud Sweden Hackathon Winner. Natural language search for retail buyers using Superlinked semantic search.