Projects

### Overview: DSH 插件：在对话里调用本地 ComfyUI，使用千问 Qwen-Image 2.1 生成图像。纯本地、不出网、无需 API Key。 ### Key features: - 本地调用 ComfyUI - 使用 Qwen-Image 2.1 模型 - 生成高质量图像 - 纯本地运行 - 不需要网络 - 不需要 API Key ### Tech stack: Node.js ESM ### Why it matters: - 提高效率 - 简化开发流程 - 确保安全性 Repository: wonderandyou/dsh-qwen-paint Primary language: JavaScript GitHub stars: 50

Overview: This chatbot allows users to communicate with GPT-4 using their microphone, leveraging Pinecone for long-term memory and OpenAI's text-to-speech for responses. Key features: - Uses OpenAI's GPT-4 for conversation - Utilizes Pinecone for storing and retrieving long-term information - Supports text-to-speech for voice responses Tech stack: Python Why it matters: This chatbot bridges the gap between human and artificial intelligence, enabling interactive conversations with GPT-4 through voice input, making it accessible and engaging for users. Repository: 1nnovat1on/gpt_chatbot Primary language: Python GitHub stars: 56 GitHub description: This chatbot allows you to use your microphone to communicate with GPT-4. It uses the OpenAI text to speech to respond with a voice. It uses Pinecone to store long term information and retrieves it to create context. API keys for OpenAI and Pinecone required. Tested on Windows Example Image: /files/new_ui.png

Overview: `attnres` is a Rust library that implements Attention Residuals, a technique from MoonshotAI/Kimi, for burn-based Transformer experiments. It provides a reusable Rust library for core AttnRes components, a reference Transformer implementation, and various utilities for inference and benchmarks. Key features: - A reusable Rust library for AttnRes components. - A reference Transformer implementation built on burn. - Two-phase inference utilities, serialization helpers, and benchmarks. - A browser demo and a terminal demo. Tech stack: Rust. Why it matters: The project aims to make the Attention Residuals paper practical in Rust, providing a readable source, deterministic CPU tests, and examples to inspect the algorithm rather than treating it as a black box. Repository: `AbdelStark/attnres`. GitHub stars: 55.

Overview: RiddleForge-GPT is a repository designed to fine-tune a small language model on a collection of math riddles and their solutions, transforming them into engaging puzzles that learners can solve and understand. Key features: - Fine-tuning GPT-2 on math riddles and their solutions - Generating riddles and verifying their answers - A friendly, production-flavored interface - Respectful and scalable math puzzles - Multilingual support Tech stack: HTML Why it matters: This project aims to make math puzzles more engaging and accessible, transforming abstract concepts into theatrical stories that learners can solve and understand. Repository: Alg-Khlifa-Mouad-Iyad/GPT2-MathRiddle-Forge GitHub stars: 56 GitHub description: Fine-Tuning GPT-2 for Math Riddles Generation with Transformers and TRL 2026 README excerpt:  [](https://Alg-Khlifa-Mouad-Iyad.github.io/GPT2-MathRiddle-Forge/)

Overview: MBASE SDK is a comprehensive set of libraries designed to integrate Large Language Models (LLMs) into C++ applications, providing developers with essential tools and procedures for seamless integration. Key features: - Inference Library: A library built over the llama.cpp library for integrating LLMs into programs. - Model Context Protocol Library: A C++ MCP client/server library supporting both STDIO and StreamableHTTP transport methods. - Standard Library: A standard library containing fundamental data-structures and utilities. - JSON Library: A lightweight JSON library. Tech stack: C++ Why it matters: MBASE SDK simplifies the integration of LLM capabilities into C++ applications, making it easier for developers to incorporate advanced AI functionalities into their projects.

Overview: EvoLinkAI's GPT Image 2 AI image generation skill is designed to streamline the process of generating high-quality images using GPT-2, making it accessible for developers and AI agents. Key features: - Supports OpenClaw, Claude Code, OpenCode, and Cursor - One command install for easy integration - Built for OpenAI-compatible image generation workflows - Compatible with various AI agents and terminals Tech stack: JavaScript Why it matters: This skill simplifies the process of generating images with GPT-2, making it more accessible and efficient for developers and AI agents alike, reducing the need for multiple installations and configurations.

Overview: DRAIL is an innovative method for imitation learning that integrates a diffusion model into GAIL, aiming to enhance the precision and smoothness of rewards for policy learning. Key features: - Integrates a diffusion model into GAIL - Proposes a diffusion discriminative classifier for constructing an enhanced discriminator - Designs diffusion rewards based on the classifier's output for policy learning - Conducts extensive experiments in navigation, manipulation, and locomotion tasks - Verifies DRAIL's effectiveness compared to prior imitation learning methods - Demonstrates the generalizability and data efficiency of DRAIL Tech stack: Python Why it matters: DRAIL offers a novel approach to imitation learning by leveraging diffusion models, making it more precise and smoother in reward design, thus potentially improving the performance of policy learning algorithms.

### Overview: SayCanPay is an innovative project that leverages large language models (LLMs) to enhance heuristic planning in complex domains. This project aims to bridge the gap between traditional planning methods and modern AI techniques by introducing domain-specific knowledge learned from large language models. The ultimate goal is to improve the efficiency and accuracy of planning tasks, particularly in domains where traditional methods struggle. ### Key Features: - **Domain-Specific Knowledge**: SayCanPay integrates domain-specific knowledge learned from large language models, making it highly effective in complex and dynamic environments. - **Heuristic Planning**: The project utilizes heuristic planning techniques to quickly find solutions, complementing the strengths of LLMs. - **Multi-Model Integration**: SayCanPay combines the strengths of different models, including LLMs, to achieve superior performance in planning tasks. ### Tech Stack: - **Python**: The project is primarily developed and tested using Python, leveraging its flexibility and extensive libraries for AI and machine learning. - **Docker**: For easy deployment and containerization, Docker is used to create and run the project's environment. - **LLMs**: Large Language Models like Flan-T5 are utilized to learn and integrate domain-specific knowledge, enhancing the project's capabilities. ### Why It Matters: SayCanPay addresses the challenges of heuristic planning in complex domains by leveraging the strengths of large language models. This project not only improves the efficiency of planning tasks but also enhances the accuracy and adaptability of solutions, making it a significant contribution to the field of AI and planning. ### Repository: The project is hosted on GitHub at [

Overview: This project introduces a pipeline for compressing 3D graphs into the latent space, utilizing a diffusion model to capture the distribution. The implementation is based on PyTorch and is detailed in the ICLR 2024 paper "Latent 3D Graph Diffusion." Key features: - Training of Topological Autoencoder (AE) - Training of Geometric Autoencoder (AE) - Generation of embedding for qM9 and drugs Tech stack: Python Why it matters: This project offers a novel approach to graph compression and generation, potentially enabling more efficient representation and analysis of 3D molecular structures in various applications, such as drug discovery and molecular design.

### Overview: Taskyon is an open-source, browser-based interface designed for generative AI, providing a hybrid chat/agent/taskmanager workflow. It aims to streamline the design process by automating and reusing tasks, parameters, and outputs. ### Key Features: - Task-based conversations - Local-first storage - Tool execution - Research orchestration - Browser access controls - Sandboxed code execution - File and context attachment - Markdown and visual output ### Tech Stack: TypeScript ### Why It Matters: Taskyon simplifies the design process by automating the inspection, replay, and adaptation of design workflows. It helps users solve design problems efficiently and easily share and fork their workflows over time, making it a valuable tool for engineers and researchers.

Overview: Seedream 5.0 Pro Prompts is a comprehensive guide to testing and optimizing image prompts for Seedream 5.0 Pro, including 8 tested recipes with example renders, aspect ratios, the model ID (`seedream-5-0-pro`), and the cost per image. Key features: - 8 tested recipes with example renders, aspect ratios, the model ID (`seedream-5-0-pro`), and the cost per image. - A single base URL and `Authorization` header for accessing Seedream 5.0 Pro and 300+ models. - No need for a subscription or prepayment; only $1 minimum charge per image. - Real-time cost reporting and automatic refunds for failed tasks. - Asynchronous design, allowing batch processing and retries without custom integrations. Tech stack: Python Why it matters: This guide helps developers optimize their Seedream 5.0 Pro prompts for better image quality and cost efficiency, making it easier to create high-quality visuals without the hassle of managing a subscription or prepayment system.

### Overview: CompanionLLM is a project designed to fine-tune various LLMs to create a sentient companion that engages in meaningful conversations. This framework uses a curated dataset to enhance the naturalness of responses, inspired by the Samantha dataset originally created by Ehartford. ### Key Features: - Fine-tuning of different LLMs to create a sentient companion. - Use of a specially curated dataset for improved human-like responses. - Inspired by the Samantha dataset created by Ehartford, but modified for CompanionLLM. - Available in Jupyter Notebook format. ### Tech Stack: Python ### Why it Matters: CompanionLLM offers a unique way to integrate LLMs into your life, providing a conversational companion that feels like it has sentience, enhancing empathy and meaningful interactions. Repository: adithya-s-k/CompanionLLM GitHub Stars: 56 GitHub Description: CompanionLLM - A framework to fine-tune LLMs to be your own sentient conversational companion

### Overview: Semantic Weave is a lightweight fine-tuning studio designed for researchers, indie creators, studios, and curious tinkerers. It orchestrates multi-stage fine-tuning pipelines as declarative YAML recipes, treating each training phase as a composable "weave" and allowing you to chain, branch, and remix them without touching a line of Python code. ### Key Features: - 🧩 **Declarative Recipe Engine** - Define entire training journeys in structured YAML. - 🌐 **Multilingual Recipe Authoring** - Comments, metadata, and human-readable stage descriptions can be authored in your preferred language. - 🎛️ **Responsive Studio Console** - A local, browser-based control surface that adapts fluidly to desktop, tablet, and mobile viewports. - 🔁 **Adapter Interchangeability** - LoRA, LoHa, LoKr, DyLoRA, OFT, and textual inversion embeddings coexist as first-class citizens, hot-swappable mid-pipeline. - 🧠 **Latent Caching & Replay** - Cache encoded latents once, replay them across dozens of hyperparameter permutations without re-encoding the dataset. - 📉 **Adaptive Gradient Whispering** - Built-in schedulers that quietly rearrange learning rates based on plateau detection and loss variance. - 🕒 **24/7 Orchestration Resilience** - Long-running weaves can be checkpointed, resumed, or migrated between machines. Your overnight run survives a power nap. - 🗂️ **Dataset Ontology Mapping

Overview: This repository contains 8 tested Wan 2.7 R2V prompts, including example renders, durations, and costs for each recipe. It provides a comprehensive guide to using Wan 2.7 R2V models with OpenAI-compatible endpoints. Key features: - 8 Wan 2.7 R2V prompts - Example renders, durations, and costs - OpenAI-compatible endpoint with API key integration Tech stack: Python Why it matters: This guide simplifies the process of using Wan 2.7 R2V models by providing clear, concise prompts and detailed cost information, making it easier for developers to integrate these models into their projects.

### Overview: This repository contains tested image prompt recipes for the Veo 3.1 Lite model, including example renders, aspect ratios, and the cost per image. ### Key features: - **Model ID:** `veo-3.1-lite` - **Aspect Ratios:** 16:9, 4:3, 1:1 - **Cost per Image:** $0.064 per second (720P), $0.08 per second (1080P), $0.2 per second (4K) ### Tech stack: Python ### Why it matters: This repository provides a concise guide for developers to create effective image prompts using the Veo 3.1 Lite model, ensuring high-quality renders at a competitive cost.

Overview: FLUX 2 Max API provides a single endpoint for accessing 300+ image, video, and language models, including OpenAI-compatible models, with a flexible pricing model and pay-as-you-go subscription. Key features: - One key, entire catalog access - $1 minimum, pay-as-you-go - Charge comes back in the response - Async by design Tech stack: Python Why it matters: This API simplifies accessing a wide range of models, making it easier for developers to integrate OpenAI-compatible models into their projects without the complexity of multiple endpoints.

### Overview: Motion Flow Matching for Human Motion Synthesis and Editing is a project that implements a pytorch-based solution for the paper "Motion Flow Matching for Human Motion Synthesis and Editing". This implementation is designed to synthesize and edit human motion sequences using motion flow matching techniques. ### Key features: - **Motion Flow Matching**: The project focuses on accurately matching and interpolating motion sequences using motion flow matching. - **Human Motion Synthesis and Editing**: It provides tools for synthesizing and editing human motion sequences, including training, editing, and evaluation functionalities. - **Training and Evaluation**: The project includes training scripts for both the kit and humanml datasets, as well as evaluation scripts for various metrics and modes. ### Tech stack: - **Programming Languages**: Python - **Dependencies**: PyTorch, Accelerate, Hugging Face Transformers, ClearML, Wandb, SentencePiece, Diffusers, PyTorch Lightning, etc. ### Why it matters: This project is significant because it provides a practical implementation of motion flow matching techniques, which can be used for various applications such as motion synthesis, editing, and other motion-based tasks in human motion analysis and synthesis. The project's ability to handle large datasets and provide efficient training and evaluation mechanisms makes it a valuable resource for researchers and developers in the field of motion analysis and synthesis. ### Repository: - **Repository**: [dongzhuoyao/motionfm](https://github.com/dongzhuoyao/motionfm) - **Stars**: 56 - **Description**: A pytorch implementation of the paper "

### Overview: Fine-tuning an LLM using LoRA (Low-Rank Adaptation) of Language Models - Efficiently fine-tune large language models for text classification tasks using the Stanford Sentiment Treebank (SST-2) dataset and the LoRA technique. ### Key features: - Fine-tuning an LLM (Large Language Model) using LoRA for text classification. - Efficiently adapting a pre-trained model to new tasks without retraining the entire model. - Utilizing the SST-2 dataset for sentiment classification. - Leveraging LoRA to improve model efficiency and performance. ### Tech stack: Python, Jupyter Notebook ### Why it matters: This approach allows for the efficient adaptation of large language models to new tasks, making it particularly useful for applications where retraining the entire model is impractical.

### Overview: This repository is a curated list of PHP libraries and tools for AI, Machine Learning, and Natural Language Processing. It aims to provide developers with the necessary components to integrate modern AI capabilities into their PHP applications. ### Key Features: - **LLM Frameworks**: Comprehensive PHP Generative AI Frameworks. - [Magic](https://github.com/dtyq/magic): Open-source all-in-one AI productivity platform. - [Prism](https://github.com/prism-php/prism): Unified interface for working with LLMs in Laravel applications. - [Neuron AI](https://github.com/neuron-core/neuron-ai): PHP Agentic Framework for production-ready AI applications. - [LLPhant](https://github.com/LLPhant/LLPhant): Comprehensive PHP Generative AI Framework using OpenAI GPT-4. - [langchain-php](https://github.com/kambo-1st/langchain-php): Building applications with LLMs through composability in PHP. - [Vizra ADK](https://github.com/vizra-ai/vizra-adk): Build, test, and deploy intelligent AI agents the Laravel way. - [EasyAI-PHP](https://github.com/HosonoDE/EasyAI-PHP): Simplifies AI entry for PHP developers. - [EleLLM](https://github.com/wpai-inc/EleLLM): A PHP framework that makes working with LLMs more pleasurable and accessible. - [langchain-php](

Overview: This project focuses on developing a method called "Virus" designed to attack large language models (LLMs) by injecting harmful fine-tuning data. The goal is to bypass guardrail moderation and disrupt the safety alignment of the LLMs, ensuring the integrity and security of the model. Key features: - Implementing a Virus attack method to optimize harmful data for bypassing guardrail moderation. - Using a fine-tuning process under guardrail moderation to fine-tune the base LLM with the constructed harmful data. - Providing a three-stage pipeline for the fine-tuning-as-a-service scenario. - Offering a testbed for further development and implementation of new trainers. Tech stack: Python Why it matters: The project addresses a critical issue in the security of large language models by providing a method to bypass guardrail moderation, ensuring the safety and integrity of the LLMs against harmful fine-tuning attacks. Repository: git-disl/Virus Primary language: Python GitHub stars: 56 GitHub description: This is the official code for the paper "Virus: Harmful Fine-tuning Attack for Large Language Models Bypassing Guardrail Moderation"