Project directory

Projects

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

Shell 8 stars

EvoLinkAI/OpenClaw-Image-Generator

A Claude Code skill (standard Skill format) for generating AI images via the Evolink Z-Image-Turbo API. Cross-platform, zero dependencies.

Overview: This is a Claude Code skill designed to generate AI images using the Evolink Z-Image-Turbo API. It is cross-platform, requiring no shell escaping or dependencies. Key features: - Cross-platform compatibility with `curl` for native support on Windows, macOS, and Linux. - No shell escaping issues, ensuring user prompts are passed directly. - Auto-download of generated images as `evolink-{timestamp}.webp`. - Configurable parameters for size, seed, and polling settings. - Standard Skill structure with progressive disclosure. Tech stack: Shell Why it matters: This skill allows for the generation of AI images through a straightforward and efficient process, leveraging the Evolink API without the need for additional dependencies or shell scripting. Repository: EvoLinkAI/OpenClaw-Image-Generator GitHub stars: 8 GitHub description: A Claude Code skill for generating images via the Evolink Z-Image-Turbo API. Cross-platform, zero dependencies.

Go 8 stars

FraiseHQ/fraise

In-memory knowledge store for AI agents

Overview: Fraise is a memory database designed specifically for AI agents, allowing them to query knowledge directly in a token-based language. Key features: - **Query Language**: Built for AI agents, allowing them to interact with the database in a language optimized for tokens rather than humans. - **In-Memory Storage**: Data is stored in memory, ensuring immediate access and high performance. - **No Infrastructure**: Runs entirely in memory, eliminating the need for any external infrastructure. Tech stack: Go. Why it matters: Fraise simplifies the interaction between AI agents and their knowledge, making it easier for them to retrieve information directly and quickly. Repository: FraiseHQ/fraise GitHub stars: 8 GitHub description: In-memory knowledge store for AI agents.

Jupyter Notebook 8 stars

GURPREETKAURJETHRA/Building-RAG-With-OpenAI-GPT-4o-omni-Model-Using-Objectbox-Vector-Database

Building RAG With OpenAI GPT-4o(omni) Model Using Objectbox Vector Database

### Overview: This project showcases the integration of OpenAI's GPT-4o model with an Objectbox Vector Database to create a highly efficient and natural RAG application. The model's ability to respond to various inputs and generate outputs in multiple formats makes it highly valuable for applications requiring comprehensive understanding and interaction. ### Key Features: - Utilizes OpenAI's GPT-4o model for creating a RAG application. - Integrates Objectbox Vector Database for efficient querying and storage of data. - Matches GPT-4o's performance on text and code, with significant improvements in non-English language understanding and faster response times. ### Tech Stack: Jupyter Notebook ### Why It Matters: GPT-4o’s ability to handle a wide range of inputs and generate outputs in multiple formats makes it a game-changer for applications requiring comprehensive understanding and interaction. Its superior performance in vision and audio understanding, along with its speed and cost-effectiveness, positions it as a leading model in natural language processing and beyond.

Python 8 stars

GeorgeVern/AFTERV1.0

Python source code for EMNLP 2020 Findings paper: "Domain Adversarial Fine-Tuning as an Effective Regularizer".

Overview: This repository contains Python source code for the EMNLP 2020 Findings paper titled "Domain Adversarial Fine-Tuning as an Effective Regularizer". The project proposes a new type of regularizer for fine-tuning pretrained language models to prevent catastrophic forgetting during the process. Key features: - Implementing a domain adversarial fine-tuning process - Using a gradient reversal layer to minimize the task-specific loss and maximize the loss of a domain classifier - Extending the standard fine-tuning process with an adversarial objective to preserve knowledge captured during pretraining Tech stack: Python Why it matters: This project demonstrates the effectiveness of domain adversarial fine-tuning as a regularizer, improving performance on various GLUE datasets compared to standard fine-tuning methods.

TypeScript 8 stars

GridLLM/MicroModel

Overview: MicroModel is a simple proxy server designed to forward OpenAI API requests and automatically save conversations as ChatML format, supporting per-workflow configuration. Key features: - Automatic workflow assignment based on port - Support for multiple workflows with different configurations - Saves conversations in ChatML format with automatic date-based organization Tech stack: TypeScript Why it matters: MicroModel simplifies the process of integrating OpenAI's API with a per-workflow configuration, making it easier to manage and analyze conversations for various workflows. Repository: GridLLM/MicroModel Primary language: TypeScript GitHub stars: 8 GitHub description: No description provided. README excerpt: # MicroModel A simple proxy server that forwards OpenAI API requests and automatically saves conversations as training data in ChatML format. Now supports per-workflow configuration with individual API hosts, keys, and ports. > [!NOTE] > Supports both `/v1/completions` and `/v1/chat/completions` endpoints. ## Quick Start 1. Clone and install: ```bash git clone https://github.com/GridLLM/MicroModel.git cd MicroModel npm install ``` 2. Create `config.json` file (required): ```json { "default": { "OPENAI_API_HOST": "http://localhost:11434", "API_KEY": "", "PORT": 3001 }, "custom_workflow": { "OPENAI_API_HOST": "http://localhost:11434",

Python 8 stars

Intelligent-Microsystems-Lab/SplitZip

SplitZip: Ultra Fast Lossless KV Compression for Disaggregated LLM Serving

Overview: SplitZip is a GPU-friendly lossless compressor designed specifically for KV cache transfer in disaggregated LLM serving. It preserves BF16 KV tensors bitwise while reducing transfer volume and maintains both compression and decompression on the latency-critical GPU path. Key features: - Lossless BF16 KV cache compression with bitwise round-trip recovery. - Offline-calibrated, Top-16 exponent codebooks with chunk-local sparse escape metadata. - GPU encode and decode kernels implemented in Triton. - Reproduction scripts for WikiText-2/Qwen3-32B exponent statistics and codec-path throughput measurements. Tech stack: Python Why it matters: SplitZip is optimized for GPU performance and provides a significant speedup over general-purpose compressors, making it ideal for real-time KV cache transfer in disaggregated LLM serving environments.

Swift 8 stars

JadenGeller/similarity-topology

Efficient nearest neighbor search in Swift

Overview: `similarity-topology` is a Swift library designed for efficient nearest neighbor searches in high-dimensional data using the Hierarchical Navigable Small World (HNSW) index. Key features: - HNSW Index: A multi-layer graph index for fast nearest neighbor searches. - Modules: Includes `HNSWAlgorithm` and `HNSWDurable` for both generic and durable storage implementations. Tech stack: Swift Why it matters: The HNSW index is crucial for applications requiring fast and efficient nearest neighbor searches in high-dimensional spaces, making it an essential tool for developers working with large datasets.

JavaScript 8 stars

JuneLearn/dsh-image2-draw

DeepSeek Harness Image2 生图插件，通过第三方 OpenAI Images 兼容接口调用 gpt-image-2，只需配置 baseURL 和 API Key。 | Image2 generation plugin for DeepSeek Harness via third-party OpenAI Images-compatible APIs.

### Overview: DeepSeek Harness Image2 生图插件，通过第三方 OpenAI Images 兼容接口调用 gpt-image-2，只需配置 baseURL 和 API Key。此插件允许开发者在 DeepSeek Harness 中生成图像，无需直接在 Harness 中生成图像。 ### Key features: - **Image2 Draw**: 添加一个“Image2 Draw”卡片到“Settings > Plugins > Configurable plugins”中。 - **仅需配置**：只需提供一个 `baseURL` 和一个 `API Key`，默认模型为 `gpt-image-2`。 - **自动后缀**：接受一个短的 `baseURL`，如 `https://example.com/v1`，并自动添加 `/images/generations`。 - **编辑端点**：生成图像编辑端点为 `/images/edits`，可选地通过 `editURL` 进行重写。 - **文本到图像**：提供 `image2-generate` 用于 1-8 个文本到图像的生成。 - **图像到图像**：提供 `image2-edit`，支持 1-8 个 PNG、JPEG 或 WebP 参考图像。 - **自适应尺寸**：支持适应性横竖屏和验证后的自定义尺寸。 - **直接显示**：在对话中直接显示生成的图像，使用专用工具卡片显示缩略图、放大和保存为操作。 - **无注入**：在客户端卡片中解析图像附件，不将图像内容注入

C++ 8 stars

Just1truc/kernel_lens

Deploy any research model right away

### Overview: KernelLens is a tool that compiles PyTorch models with custom `@triton.jit` kernels into native C++ artifacts for ONNX Runtime and NVIDIA TensorRT. It simplifies the process by generating zero-copy VRAM binding, enabling dynamic grids, and supports multiple TensorRT versions. ### Key features: - **Dual-pass tracing**: Captures PTX, shared memory, and strides; reconstructs dynamic launch grids. - **Automated plugin synthesis**: Generates `OrtCustomOp` and `IPluginV2DynamicExt` sources, compiling with `nvcc`/`g++`. - **Dynamic grids in C++**: Launch-grid expressions transpiled to C++ integer math, serving dynamic batch and sequence lengths. - **One package, several TensorRT versions**: Interface choice from environment, TensorRT 10+ emits `IPluginV3` plugins, TensorRT 11+ emits `IPluginV2DynamicExt` plugins. - **Shape-generic engines**: Leading dimension as a symbolic axis, serving any size in a declared range via a TensorRT optimization profile. - **Zero-copy VRAM binding**: PyTorch `data_ptr()` values registered directly with each engine, no host staging copies. - **Blackwell PTX ISA clamping**: Triton-generated `.version 9.3` header rewritten to `.version 9.0` for CUDA driver JIT. - **FP8 & sub-byte INT4**: 1-byte pointer mapping and dtype-accurate in-place copies. - **Cold-start detection**: Detect

Go 8 stars

Kardbord/gopenai

Unofficial Go (Golang) bindings for the OpenAI API.

### Overview: This is an unofficial Go (Golang) bindings for the OpenAI API, providing comprehensive support for various functionalities including audio, chat, completions, embeddings, files, fine-tuning, images, models, and moderations. ### Key Features: - Audio - Chat - Completions - Embeddings - Files - Fine-Tuning - Images - Models - Moderations ### Tech Stack: Go ### Why it Matters: This library is essential for developers looking to integrate OpenAI's capabilities into their Go projects, offering a straightforward and efficient way to interact with the API. It is maintained by the community and is not affiliated with OpenAI, making it a valuable resource for those seeking to leverage the API without official support.

Python 8 stars

LukaVuli/Entity_Neutering

A methodology to pre-process text data for preventing lookahead bias in Large Language Models (LLMs), particularly in financial text analysis.

Overview: Entity Neutering is a methodology designed to pre-process text data to prevent lookahead bias in Large Language Models (LLMs), particularly in financial text analysis. It anonymizes financial documents and news articles by masking and/or paraphrasing identifying features, ensuring that model predictions are based solely on the textual content. Key features: - Iterative masking and paraphrasing until the same LLM cannot infer the entity. - Optional LLM-based masking and paraphrasing. - Flexible processing modes: masking only, paraphrasing only, or complete neutering. - Up to 8 rounds of refinement. - Built-in identification testing to verify neutering success. - Robustness testing with Ollama models. Tech stack: Python. Why it matters: This methodology helps ensure that LLMs do not use prior knowledge about specific companies, industries, or time periods, thereby preventing lookahead bias and improving the fairness and reliability of financial text analysis.

C++ 8 stars

MarmotTech/FlexInfer

Artifact code for FlexInfer, a flexible offloading system for on-device LLM inference built on llama.cpp.

### Overview: FlexInfer is a mobile-oriented LLM inference runtime designed to run large quantized models under tight memory budgets. It extends the `llama.cpp` and `ggml` stack with prefetch-aware execution paths, memory planning, and Linux host plus Android experiment scripts. ### Key features: - Prefetch-aware execution paths - Memory planning - Linux host plus Android experiment scripts ### Tech stack: C++ ### Why it matters: FlexInfer provides a flexible and efficient way to run large quantized models on mobile devices, optimizing memory usage and performance. It enables developers to leverage powerful LLMs without the constraints of high memory budgets, making it suitable for a wide range of applications.

Python 8 stars

Mizistein/omlx

🤖 Optimize LLM inference on Mac with continuous batching and SSD caching managed from your menu bar for efficient performance.

Overview: omlx is an application designed to optimize large language model (LLM) inference on Apple Silicon Macs, providing efficient performance through continuous batching, SSD caching, and seamless menu bar control. Key features: - Continuous batching - SSD caching - Menu bar control - OpenAI API compatibility - Lightweight system requirements Tech stack: Python Why it matters: omlx simplifies the use of LLMs on Macs, making large language model inference faster and more efficient, while managing resources effectively to ensure smooth performance. Repository: Mizistein/omlx Primary language: Python GitHub stars: 8 GitHub description: 🤖 Optimize LLM inference on Mac with continuous batching and SSD caching managed from your menu bar for efficient performance.

Python 8 stars

MukulRay1603/Irminsul

Fine-tuned Llama 3.1 8B · QLoRA · RAG serving stack · FastAPI · Pinecone · Azure

### Overview: Irminsul is a Genshin Impact AI assistant that answers questions based on game data, refusing AI-generated content and leaks. It uses a combination of QLoRA, Pinecone, and FastAPI to serve answers in multiple languages, ensuring source-cited responses and a leak-gated approach. ### Key Features: - **Answers from Game Data:** Every answer links to its source, and game data is marked as *No AI used*. - **Leak-Gated:** Questions that cannot be proven released before indexing are refused. - **Builds and Teams:** Answers are checked against game data, with newer characters using quick guides or stating that no build exists yet. - **Quests and Lore:** Answers are dated to prevent story spoilers. - **Live Game News:** Banner updates, event information, and expired codes from HoYoverse. - **Spoiler-Free Lore:** Answers stop at the player's Archon Quest progress. - **Supports Multiple Languages:** Answers are provided in Chinese, Japanese, Korean, Spanish, Portuguese, and more. - **Interactive Tools:** Answers can be checked, and patch notes are provided by version. - **Daily Farming Planner:** Saved on the player's device. ### Tech Stack: - Python ### Why It Matters: Irminsul provides accurate, source-cited answers to Genshin Impact questions, ensuring fair use without requiring an account and avoiding leaks or story spoilers.

Jupyter Notebook 8 stars

NisugaJ/rag-test

An Advanced RAG Pipeline

Overview: This project is an advanced RAG (Retrieval-Augmented Generation) pipeline designed to enhance text generation capabilities by integrating knowledge from multiple sources. Key features: - Integration of multiple sources for enhanced knowledge retrieval - Advanced text generation models - Customizable pipelines for various applications Tech stack: Jupyter Notebook Why it matters: This project offers a powerful tool for researchers and developers to improve the quality and relevance of their text generation tasks by leveraging diverse and accurate knowledge sources.

Python 8 stars

Notnaton/microllm

My own implementation to run inference on local LLM models

Overview: A lightweight implementation for running inference on local Large Language Models (LLMs). Key features: - Read and parse `.gguf` files - Simplified and optimized token generation Tech stack: Python Why it matters: This project provides a straightforward and efficient way to run inference on local hardware, making it useful for developers and researchers working with LLMs.

Python 8 stars

OctipusPrime/gmGPT

LangChain application acting like a Game Master for a table top rpg of your choice.

### Overview: A LangChain application designed to act as a Game Master for tabletop RPGs, integrating various fan-favorite settings and characters. ### Key Features: - **GUI Interface**: A user-friendly graphical interface for setting up a game, creating characters, and starting the adventure. - **CLI Commands**: Command-line interface for more advanced users, allowing for easier navigation and management of the game. - **Vector Stores**: Integration with Qdrant for efficient storage and retrieval of game-related data. - **Adaptive Memory**: The ability to remember game progress and context beyond the initial setting. ### Tech Stack: Python ### Why it Matters: This application streamlines the process of creating and managing a tabletop RPG, making it accessible and efficient for players and GMs alike.

Python 8 stars

Orion-zhen/transAPI

OpenAI compatible API purely based on Transformers

Overview: TransAPI is an OpenAI compatible API purely based on Transformers, offering an alternative to unsupported models and models not available through GGUF. Key features: - OpenAI compatible API - Native integration with Transformers - Multimodal support (Text-to-Text and Image-Text-to-Text) - Reverse proxy support (e.g. Nginx) - Multiple sampling strategies including beam search - `torch.compile` for faster inference - `bitsandbytes` in-flight quantization - Various quantization formats supported by Transformers - Concurrent inference with asyncio - Streaming and non-streaming responses - Smart KV Cache offloading Tech stack: Python Why it matters: This project provides an alternative for unsupported models and models not available through GGUF, offering an OpenAI compatible API based entirely on Transformers, allowing you to run and test any model compatible with Transformers. Repository: Orion-zhen/transAPI Primary language: Python GitHub stars: 8 GitHub description: OpenAI compatible API purely based on Transformers

Python 8 stars

PanGan21/tunebench

A CLI for comparing full fine-tuning, LoRA, and layer freezing on small LLMs.

Overview: Tunebench is a CLI tool designed for developers to compare full fine-tuning, LoRA, and layer freezing techniques on small language models. Key features: - Full fine-tuning - LoRA (Layer-wise Adaptive Rate) - Layer freezing - Weight difference analysis - Catastrophic forgetting check - Gradient norm tracking - Layer-wise learning rate decay Tech stack: Python Why it matters: Tunebench simplifies the process of evaluating different fine-tuning strategies on small language models, making it easier for developers to understand and optimize their models. Repository: PanGan21/tunebench GitHub stars: 8 GitHub description: A CLI for comparing full fine-tuning, LoRA, and layer freezing on small LLMs.

Python 8 stars

PierreMarieCurie/train-rf-detr-OIDv7

One‑line fine‑tuning of RF‑DETR on selected classes from OpenImages V7

### Overview: This repository provides a streamlined method to fine-tune the RF-DETR model on custom classes from the Open Images V7 dataset, making it accessible for developers looking to integrate object detection capabilities into their projects. ### Key Features: - Fine-tuning of RF-DETR on selected classes from Open Images V7 - Minimal setup required, leveraging Python and the UV package - Easy to use with a single command for fine-tuning - Supports fine-tuning on multiple classes simultaneously - Inference capabilities for new images ### Tech Stack: Python 3.10+ and UV package ### Why it Matters: This approach simplifies the process of integrating advanced object detection models into projects, making it accessible to developers without extensive machine learning expertise.