Projects

Overview: ToolBrain is a Python library designed for training agentic systems using reinforcement learning (RL). It supports various algorithms like GRPO, DPO, and supervised learning, flexible reward functions, and efficient tool management. Key features: - 🤬 Learning algorithms: GRPO, DPO, and supervised learning support. - 🎯 Flexible rewards: Custom reward functions or LLM-as-judge. - 🔧 Tool management: Scalable retrieval for large tool collections. - 📊 Knowledge distillation: Distill large teacher models into smaller student models. - 🚀 Zero-learn: Automatically generate training tasks. - 🧠 Efficient training: Supports FP16 finetuning, LoRA, Unsloth, and BitsAndBytes for resource-efficient training. - 🧠 Multiple agent frameworks: Supports SmolAgent and LangChain, with more coming soon. Tech stack: Python 3.10+. Why it matters: ToolBrain simplifies the process of training agentic systems with reinforcement learning, making it accessible for developers and researchers alike. Repository: ToolBrain/ToolBrain GitHub stars: 206 GitHub description: A framework for agentic tool use training with reinforcement learning. README excerpt: [Link to README](https://github.com/ToolBrain/ToolBrain/blob/main/README.md) --- # ToolBrain 🧠 ### *RL Framework for Agentic Tools* ToolBrain is a lightweight open-source Python library for training **agentic systems** with effective tool usage and built-in reinforcement learning. 📚 **Website:**

### Overview: Gemini API is an unofficial Python wrapper that allows users to interact with Google Gemini through reverse-engineering, providing a workaround for frequent authentication issues or inability to use Google Authentication. This project is a collaboration with Antonio Cheong. ### Key Features: - Installation - Authentication - Quick Start - Usage: Initialization, Generate Content, Send Request, Text Generation, Image Generation, Retrieving Images, Fix Context Setting, Changing Selected Response ### Tech Stack: Python ### Why it Matters: This package offers a solution for developers facing authentication issues with Google Gemini, providing a practical workaround for frequent login failures.

Overview: Fusilli is a Python package designed to handle multi-modal data fusion in machine learning. It provides a comprehensive platform for combining different types of data, including tabular and image modalities, for various prediction tasks such as regression, binary classification, and multi-class classification. Key features: - Support for Tabular-Tabular and Tabular-Image fusion scenarios - Integration with PyTorch Lightning and PyTorch Geometric - Compatibility with regression, binary classification, and multi-class classification tasks - Visualization of real values vs. predicted values for easy evaluation Tech stack: Python, PyTorch Lightning, PyTorch Geometric Why it matters: Fusilli simplifies the process of multi-modal data fusion, making it accessible to developers and researchers without requiring extensive knowledge of deep learning frameworks. Its modular design and comprehensive documentation make it an ideal tool for anyone working on multi-modal data analysis and machine learning projects.

### Overview: Strata NVFP4 is a fork of the original Strata project, featuring Qwen3.8-Flash-Next with 125B MoE and NVIDIA NVFP4 experts. This setup runs on one RTX 20-50 card with 12 GB of VRAM and 64 GB of RAM, and includes Hugging Face's Qwen in NVFP4 format. The models are highly accurate and efficient, with a speed of 32K tokens/s on an RTX 5090 with 128 GB of RAM. ### Key features: - **Qwen3.8-Flash-Next with 125B MoE** - **NVIDIA NVFP4 Experts** - **GPU: RTX 20-50, 12 GB VRAM, 64 GB RAM** - **Hugging Face's Qwen in NVFP4 Format** - **Accuracy: 95.0% against Q8_0** - **Speed: 32K tokens/s on an RTX 5090 with 128 GB of RAM** - **Setup Downloads Models from Hugging Face** - **Models:** - Q8_0: 100% accuracy - NVFP4 + Q8_0 mix: 95.0% - Unsloth UD-Q4_K_XL: 94.1% - NVIDIA NVFP4: 93.6% -

Overview: This project is an AI-driven adult video generator that uses text-to-image diffusion and motion to create uncensored, 4K NSFW videos from text prompts. It offers a unique blend of text-to-image and motion generation, making it a powerful tool for creating adult content. Key features: - Text-to-Image Diffusion - Motion Generation - 4K Output - NSFW Content - Uncensored Model Presets Tech stack: Python Why it matters: This project is significant as it combines cutting-edge AI techniques to generate adult content, offering a fresh approach to the adult video generation industry. The uncensored model presets and 4K output make it a valuable tool for both creators and consumers of adult content. Repository: wlnhnoqefoxt/adult-ai-video Primary language: Python GitHub stars: 206

### Overview: Leveraging continuous, contradictory project interaction logs, Dynamics-memory creates a bounded, self-correcting long-term memory layer for LLM agents. This innovative approach distills complex, evolving project interactions into a structured, updatable memory that can adapt and correct itself as it learns. ### Key Features: - **Continuous Log Integration**: Seamlessly integrates and processes continuous project interaction logs. - **Self-Correction Mechanism**: Automatically corrects errors and inconsistencies in the memory. - **Memory Boundedness**: Ensures the memory is limited in size and can be updated as needed. - **Value Dynamics**: Tracks and adjusts the value of projects based on their interactions. - **Tension Adjudication**: Handles and resolves conflicts between projects dynamically. - **Causal-Replay Evaluations**: Evaluates the causal relationships and replay history of interactions. ### Tech Stack: TypeScript ### Why It Matters: This project addresses the need for robust, adaptive memory systems that can learn from and adapt to complex project interactions. By integrating continuous project logs and implementing a self-correcting mechanism, Dynamics-memory provides a valuable tool for LLM agents to improve their decision-making and project management capabilities.

Overview: AutoAgents is a project focused on complex question answering using Large Language Models (LLMs), enhancing reasoning and information-seeking capabilities. Key features: - Enhanced LLM reasoning - Improved information-seeking abilities - Streamlit-based application development Tech stack: Python Why it matters: AutoAgents showcases advancements in LLMs by integrating advanced reasoning and information-seeking features, making it a valuable tool for developers looking to improve their AI applications.

Overview: This repository contains the code and artifacts used to reproduce the experiments described in the paper "Reproducible scaling laws for contrastive language-image learning" by Mehdi Cherti et al. The repository includes pre-trained models, intermediate checkpoints, and datasets used in the study, as well as pointers to other related resources. Key features: - Pre-trained models across various scales - Intermediate checkpoints for training dynamics - OpenCLIP datasets and instructions for reproducibility - Scripts for downloading pre-trained models Tech stack: Jupyter Notebook Why it matters: This repository provides a reproducible framework for researchers to study and apply the scaling laws discussed in the paper, facilitating further research and development in contrastive language-image learning. GitHub stars: 205 Repository: LAION-AI/scaling-laws-openclip

### Overview: This is a study group focused on applying Machine Learning to the Audio domain, covering topics such as generating synthetic sound, transcribing audio signals, removing noise, separating audio sources, and identifying speaker identities. ### Key Features: - **Presentations and Discussions:** Explore the field of applying Machine Learning in the Audio domain. - **Resources:** Read/watch related resources before each session. - **Presentations:** Short (~10-15 mins) presentations on the topic. - **Discussions:** Keep discussing and asking questions about the topic. - **Schedule:** Topics include Intro to Audio, Automatic Speech Recognition, Text to Speech, and more. ### Tech Stack: Jupyter Notebook ### Why it Matters: This study group aims to provide a comprehensive understanding of how Machine Learning can be applied to audio processing, enabling applications such as conversational assistants, speech recognition, noise reduction, speaker identification, and more. Joining this community effort will allow you to contribute to and learn from others in the field. ### Repository: Vaibhavs10/ml-with-audio ### Primary Language: Jupyter Notebook ### GitHub Stars: 205 ### GitHub Description: Hugging Face's Machine Learning for Audio Study Group ### Schedule: - **Dec 14, 2021:** Kickoff + Overview of Audio related usecases - **Dec 21, 2021:** Intro to Audio & Automatic Speech Recognition Deep Dive - **Jan 4, 2022:** Text to Speech

Overview: CaveAgent is a stateful runtime management system designed for LLM agents, enabling them to manipulate and retrieve Python objects across turns without serialization. Key features: - Stateful Runtime Management - Variable Injection and Retrieval - Persistent Python Objects Tech stack: Python Why it matters: CaveAgent transforms LLMs into stateful runtime operators, allowing for seamless manipulation and retrieval of Python objects, breaking the traditional text-in-text-out paradigm. Repository: acodercat/cave-agent GitHub stars: 205

Overview: Mengram is a human-like memory system designed for AI agents, providing semantic, episodic, and procedural knowledge. It learns from failures and offers a free API and Python/JS SDKs, integrating with LangChain, CrewAI, and OpenClaw. Key features: - Semantic knowledge - Episodic memory - Procedural learning from failures - Free API and SDKs - LangChain, CrewAI, and OpenClaw integrations Tech stack: Python Why it matters: Mengram's ability to learn from failures and provide a human-like memory system makes it valuable for AI agents, enhancing their decision-making and problem-solving capabilities. Its free API and SDKs, combined with integrations with popular AI tools, make it accessible and powerful for developers. Repository: alibaizhanov/mengram GitHub stars: 204 GitHub description: Human-like memory for AI agents — semantic, episodic & procedural. Experience-driven procedures that learn from failures. Free API, Python & JS SDKs, LangChain, CrewAI & OpenClaw integrations.

### Overview: OpenFill is a generative tool designed for PhotoCraft and GIMP 3, allowing users to remove objects, fill selections, and expand the canvas directly within their photo editing software. It leverages local GPU processing for efficiency and runs seamlessly on Windows and Linux, including support for AMD graphics cards. ### Key Features: - **Generative Fill:** Removes objects from a selection. - **Generative Expand:** Paints new borders around the canvas edges. - **Object Removal:** Erases selected objects without prompts. - **Create New Image/Layer:** Generates a new image or layer based on a prompt. ### Tech Stack: - Made with Python and ComfyUI. ### Why It Matters: OpenFill provides a familiar workflow for users transitioning from Photoshop, allowing them to maintain their existing skills while benefiting from the advanced AI capabilities of PhotoCraft and GIMP. It runs locally on your GPU, ensuring no external dependencies or additional software are required, making it a practical solution for photo editing. ### GitHub Metadata: - Repository: `beckettharriman/openfill` - Primary Language: Python - GitHub Stars: 174 - GitHub Description: Generative Fill, Remove object, Generative Expand and Text to Image for PhotoCraft and GIMP 3. Free, local, on your own GPU through ComfyUI.

**Overview:** This repository is a comprehensive collection of resources and papers related to large language models (LLMs), focusing on the latest trends and advancements in the field. It includes papers on model training, optimization, and application, as well as resources like courses, figures, demos, and blog posts. **Key features:** - Model training and optimization - Application and LLM+ - Principle analysis - Technical improvements, such as generation, prompt engineering, and evaluation metrics - Survey and data sets **Tech stack:** Python, Markdown, LaTeX **Why it matters:** This library helps developers stay updated with the latest research and trends in LLMs, enabling them to better understand and apply these models in their projects.

Overview: This course provides a comprehensive exploration of AI engineering, focusing on Multimodal Retrieval Augmented Generation (RAG) and Amazon Nova. Learn to implement advanced RAG techniques, use LangChain with Amazon Bedrock, create multimodal AI systems, and build practical applications using AWS services. Key features: - Implement advanced RAG techniques with LangChain - Use Amazon Bedrock for foundational models - Process text and image data with advanced LLMs - Understand Amazon Nova for multimodal AI systems - Implement Multimodal RAG applications with AWS services - Evaluate and optimize AI model performance for production environments - Build an end-to-end application with Amazon Bedrock Agents and Knowledge Bases Tech stack: Python, Jupyter Notebook, AWS (Amazon Bedrock, Amazon Nova, LangChain, Amazon DynamoDB, Amazon Lambda, Amazon OpenSearch, Amazon Bedrock Agents, Amazon Knowledge Base) Why it matters: This course equips learners with practical skills to integrate AWS services into AI systems, enabling them to build context-aware AI solutions that enhance decision-making and operational efficiency.

Overview: This repository is a playful exploration of the Stable Diffusion model, allowing users to generate and manipulate images with reproducibility. The images are saved with metadata and latent information, ensuring each image can be recreated. Key features: - Generate images - Interpolate between generated images Tech stack: Python Why it matters: This project demonstrates the reproducibility of image generation, making it useful for artistic and creative projects where precise control over the image creation process is desired. Repository: gordicaleksa/stable_diffusion_playground GitHub stars: 205

### Overview: This repository provides an official PyTorch implementation of the paper "Massive Activations in Large Language Models". The repository includes code for visualizing hidden states, collecting internal activation and attention statistics, and reproducing experiments on various Large Language Models (LLMs) and Vision Transformers (ViTs). ### Key features: - Code for visualizing hidden state features on the residual stream. - Code for collecting and visualizing per-layer top activation magnitudes. - Code for running intervention analysis to zero out massive activations in specific layers. - Code for visualizing attention logits in specific layers of ViTs. ### Tech stack: Python ### Why it matters: This repository provides a comprehensive toolset for studying and visualizing the activation magnitudes and attention statistics in Large Language Models and Vision Transformers, which can help researchers better understand and potentially improve these models.

Overview: GuoFeng Webnovel is a multilingual corpus of web fiction, focusing on discourse-level and rich linguistic and cultural phenomena. It includes 14 genres and is used for research in machine translation and artificial general intelligence. Key features: - Rich linguistic and cultural phenomena - Long-range context - Discourse-level datasets for Chinese→German and Chinese→Russian - Sentence-level alignment for Chinese→English Tech stack: Ruby Why it matters: This dataset will advance existing research in machine translation and inspire novel studies with Large Language Models. Repository: longyuewangdcu/GuoFeng-Webnovel GitHub stars: 205 GitHub description: Multilingual Corpus of Web Fiction

### Overview: IMAGHarmony is a framework designed for controllable image editing, ensuring consistent object quantity and layout in multi-object scenes. It leverages a structure-aware framework and harmony-aware attention to achieve precise, stable, and semantically aligned generation. ### Key Features: - **Structure-Aware Framework (QL-Edit):** Ensures precise control over object count, category, and arrangement. - **Harmony-Aware Module (HA):** Models object structure and semantics to stabilize generation. - **Preference-Guided Noise Selection (PNS):** Selects semantically aligned initial noise to stabilize the generation process. ### Tech Stack: - Python 3.8+ - PyTorch 2.0.0+ - CUDA 11.8+ ### Why It Matters: IMAGHarmony addresses the challenge of controllable image editing in complex multi-object scenes, offering precise control over object quantity and layout, and ensuring high fidelity and coherence in editing tasks. ### Repository: - **Repository:** muzishen/IMAGHarmony - **GitHub Stars:** 205 - **Primary Language:** Python - **GitHub Description:** 🧩 IMAGHarmony 🧩: Controllable image editing with consistent object quantity and layout. A structure-aware framework that ensures high fidelity and coherence in complex multi-object edits. It integrates harmony-aware attention and preference-guided noise selection to enable precise, stable, and semantically aligned generation. ### Download Models: - Models available on [Huggingface](https://huggingface.co/kkkkg

### Overview: Rekal is a command-line tool that captures your AI coding sessions at every git commit, stores them in your repo, and lets any agent recall the reasoning behind a past decision — including the approaches that were tried and rejected. Once that work merges, it travels to your teammates over the same `git push` and `git fetch` you already run — so their agents start from your reasoning, not from nothing. No server, no account, no memory service in the middle. ### Key Features: - **Session Capture**: Rekal captures every AI session at every commit. - **Git Integration**: Stores the sessions in your repo, making them accessible via `git push` and `git fetch`. - **Personalized Memory**: As you recall, the memories stay on your machine only, never synced. - **Adaptive Reasoning**: The reasoning behind every change is recalled in ~7.5K context tokens and a few seconds. - **Team Collaboration**: Team members can recall the conversation behind any change, not just the diff. - **Stop Re-deciding**: Dead-ends already ruled out stay ruled out, preventing re-proposing them. - **Scrubbed Before Storage**: Transcripts pass through secret redaction and home-path anonymization before reaching the database, ensuring privacy. ### Tech Stack: Go ### Why It Matters: Rekal is a game-changer for teams that want to improve their agentic development lifecycle. It ensures that every AI session settles decisions, stops re-deciding, and keeps team memory in git, making it easy to collaborate

### Overview: Mneme 是 DeepSeek Harness 的跨会话记忆插件，用于存储和管理 AI 代理的记忆。它通过自动去重合并、矛盾先冻结等机制确保记忆的准确性，并提供 4-way recall（关键词、向量、BM25、实体）功能，支持本地 SQLite 和 Markdown 镜像，以及一个记忆面板。Mneme 还包括 10 个模型工具、MCP 服务器和命令行工具，旨在提供一个安全、可审计、可回放的 AI 记忆系统。 ### Key features: - **自动去重合并**：后台自动去重合并记忆库中的重复项。 - **矛盾先冻结**：对记忆库中的矛盾信息进行先冻结处理。 - **4-way recall**：支持关键词、向量、BM25、实体四种检索方式。 - **本地 SQLite 镜像**：提供本地 SQLite 镜像存储。 - **Markdown 镜像**：提供可读的 Markdown 镜像存储。 - **记忆面板**：提供一个记忆面板供用户查看和管理记忆。 ### Tech stack: - **JavaScript** ### Why it matters: Mneme 是 DeepSeek Harness 的核心记忆系统，通过提供安全、可审计、可回放的 AI 记忆功能，为 AI 代理提供了强大的记忆管理能力。它确保了记忆的准确性和可追溯性，为 AI 代理的决策提供可靠的数据支持。