Projects

### Overview This repository is a Python application that leverages DuckDB for efficient similarity searches among large datasets. The application uses embedding vectors generated from OpenAI's models to find the most similar comments from a large dataset. ### Key Features - **DuckDB Backend**: Utilizes DuckDB for storage and retrieval of embedding vectors. - **Embedding Vectors**: Generated using OpenAI's models for high-quality semantic understanding. - **Similarity Search**: Finds the 10 most similar comments from a large dataset based on embedding comparisons. ### Tech Stack - Python - DuckDB - OpenAI API key ### Why It Matters This application showcases the potential of combining DuckDB with embedding vectors for efficient similarity searches, making it a valuable tool for developers looking to improve their data retrieval capabilities.

### Overview: Rubra is an enhanced collection of open-weight large language models, designed to be ideal for agentic use cases. It leverages the capabilities of top open-source LLMs, including Meta, Google, and others, to call user-defined external tools in a deterministic manner. This feature makes Rubra models highly versatile and effective for various applications. ### Key Features: - Tool-calling capability - Enhanced from top open-source LLMs - Post-training and methods for effective skill teaching - Easy-to-use inferencing projects - Compatible with popular inferencing tools like llama.cpp and vLLM ### Tech Stack: Python-based ### Why It Matters: Rubra models are optimized for practical applications, making them highly effective and efficient for tasks requiring external tool integration. Their enhanced capabilities and compatibility with popular tools ensure seamless integration and performance in real-world scenarios. ### Repository: - Repository: rubra-ai/rubra - Stars: 153 - Description: Open Weight, tool-calling LLMs - GitHub: https://github.com/rubra-ai/rubra ### Demo: - Available in Huggingface Spaces: https://huggingface.co/spaces/sanjay920/rubra-v0.1-dev - Free and requires no login ### Run Locally: - Documentation: https://docs.rubra.ai/category/serving--inferencing - Extends llama.cpp and vLLM to run Rubra models in an OpenAI-compatible tool-calling format ### Benchmark: - Full

Overview: This repository contains the code and dataset used in the paper "Large Language Model as Attributed Training Data Generator: A Tale of Diversity and Bias" presented at NeurIPS 2023 (D&B Track). It provides a framework for developing and evaluating training data generation pipelines with Large Language Models. Key features: - Code and dataset for the NeurIPS 2023 paper - Framework for training data generation pipelines - Available datasets including original train/valid/test data, generated training data, and label names - Detailed information on datasets for AG News, SST-2/IMDB, and Yelp - Commands for loading datasets from Hugging Face Dataset Hub - Code for training data generation and classifier training Tech stack: Python Why it matters: This repository is useful for researchers and developers interested in creating effective training data generation pipelines for Large Language Models, particularly in the context of diversity and bias in training data.

### Overview: This project is an implementation of WSRL, a warm-start reinforcement learning algorithm, as well as popular actor-critic RL algorithms in JAX and Flax. The code is based on the paper "Efficient Online Reinforcement Learning Fine-Tuning Need Not Retain Offline Data" and is compatible with various environments such as D4RL, adroit, and Mujoco locomotion. ### Key features: - WSRL: A warm-start reinforcement learning algorithm. - IQL: Actor-critic algorithm. - CQL: Actor-critic algorithm. - CalQL: Actor-critic algorithm. - SAC: Actor-critic algorithm. - RLPD: Actor-critic algorithm. - TD3: Actor-critic algorithm. - REDQ: Actor-critic algorithm. - IQL policy extraction supports both AWR and DDPG+BC. ### Tech stack: Python ### Why it matters: This project provides a practical solution for fine-tuning models without the need to retain offline data, making it particularly useful in scenarios where data is limited or expensive to collect. The implementation of popular RL algorithms in JAX and Flax makes it accessible for researchers and practitioners to experiment with and evaluate these methods. ### Repository: zhouzypaul/wsrl ### GitHub stars: 152 ### GitHub description: JAX implementation of WSRL and RL baselines | ICLR 2025 ### Installation: ```bash conda create -n wsrl python=3.10 -y conda activate wsrl pip install -r

### Overview This OpenClaw skill integrates Decodo's Web Scraping API into any OpenClaw-compatible AI agent or LLM pipeline, allowing agents to call these tools directly: `google_search`, `universal`, `amazon`, `amazon_search`, `youtube_subtitles`, `reddit_post`, and `reddit_subreddit`. The skill handles JavaScript rendering, bot detection, and CAPTCHA bypass out of the box, ensuring seamless integration with Decodo's residential and datacenter proxy infrastructure. ### Key Features - **Real-time Google Search results scraping** - **Universal URL scraping** - **Amazon product page parsing (by URL)** - **Amazon search (by query)** - **YouTube subtitles/transcript by video ID** - **Reddit post content by URL** - **Reddit subreddit listing by URL** - **Structured JSON or Markdown results** - **Simple CLI interface compatible with any OpenClaw agent runtime** - **Designed for scalable AI agent web scraping** - **Minimal dependencies — just Python with Requests** - **Authentication via a single Base64 token** ### Tech Stack Python ### Why It Matters - **Zero blocks & CAPTCHAs**: Decodo's proxy infrastructure handles JavaScript rendering, bot detection, and CAPTCHA bypass, ensuring seamless integration with Decodo's residential and datacenter proxy infrastructure. - **Real-time data**: Access fresh, up-to-the-minute web data directly within your AI agent's workflow. - **LLM-optimized output**: Data is returned in structured JSON or clean Markdown, making it easy for LLMs to understand and process

### Overview: State-of-Thought enables endogenous reasoning by controlling evidence access and committing reasoning when the model's internal state indicates it is ready. ### Key Features: - **Endogenous Reasoning Control:** State-of-Thought manages reasoning from within the model's internal dynamics, rather than externally. - **Evidence Selection:** It selectively retrieves evidence when it becomes useful and stops reasoning when the model's internal state indicates the answer is ready. - **Lightweight State Management:** The framework uses a compact dynamics-geometric state to control evidence access and commit reasoning. ### Tech Stack: - Python ### Why It Matters: State-of-Thought provides a more efficient and transparent way of reasoning by leveraging the model's internal dynamics, reducing the need for external control mechanisms.

### Overview InfiGUI-G1 is an official repository for the paper "InfiGUI-G1: Advancing GUI Grounding with Adaptive Exploration Policy Optimization" accepted by AAAI 2026. This project introduces AEPO, a method that combines multi-answer generation and adaptive reward to overcome exploration bottlenecks in GUI agents, enabling robust semantic alignment. ### Key Features - **Adaptive Exploration Policy Optimization (AEPO)**: A method that integrates multi-answer generation and adaptive reward to break the exploration bottleneck, enhancing semantic alignment. - **Models**: InfiGUI-G1 series, including 3B and 7B models, trained with AEPO for efficient exploration and semantic alignment. ### Tech Stack - Python ### Why It Matters InfiGUI-G1 addresses a critical challenge in GUI agents by overcoming the exploration bottleneck, enabling models to discover correct actions for difficult semantic associations. This project demonstrates the effectiveness of AEPO in enhancing GUI grounding and semantic alignment. ### Repository The primary language for this repository is Python, and it has been starred by 152 GitHub users.

### Overview: ComfyUI-Drama3 is a text-to-speech model for voice cloning, dialogue, and inline cues, designed for use with ComfyUI. It leverages 8 GB of VRAM and 32 to 48 GB of RAM, with a native Qwen3-ASR for transcription and a voice cloning feature that allows you to clone a voice from seconds of audio. ### Key Features: - **Voice Cloning:** Clone a voice from seconds of audio and reuse it without the codec. - **Dialogue:** Up to 16 speakers in one text, starting each line with `<|speaker:N|>`. - **Cues:** Inline cues such as `[laughs]`, `[whispers]`, `[sighs]`, and plain-language directions inside the text. - **Long Texts:** Read in chunks the way Fish Audio's server does, each chunk hearing the ones before it. ### Tech Stack: ComfyUI-Drama3 is built with Python and utilizes the ComfyUI framework, leveraging expertise and layering techniques to keep the experts in system RAM and the GPU's cache of used layers. ### Why It Matters: ComfyUI-Drama3 offers a powerful text-to-speech solution that is natively integrated with ComfyUI, making it ideal for voice cloning, dialogue, and inline cues. Its efficient use of VRAM and GPU resources ensures smooth performance, making it a valuable tool for developers and creators looking to enhance their applications with advanced text-to-speech capabilities.

Overview: Yo'Chameleon is a project that introduces the first attempt to study personalization for large multimodal models, specifically focusing on image generation. It leverages soft-prompt tuning to embed subject-specific information and enhance image quality in a few-shot setting. Key features: - Yo'Chameleon leverages soft-prompt tuning to embed subject-specific information. - It enhances image quality in a few-shot setting through a "soft-positive" image generation approach. - Yo'Chameleon is trained with a self-prompting optimization mechanism to balance performance across multiple modalities. Tech stack: Python Why it matters: Yo'Chameleon represents a significant step towards personalization in large multimodal models, offering a more efficient and effective way to learn concepts and encode visual attributes compared to prompting baselines.

Overview: novel-studio is an open-source AI novel generator designed for long-form fiction, offering a robust system for world-building, character development, and narrative progression. It ensures the preservation of the story's integrity and avoids the loss of context, providing a reliable and verifiable production pipeline. Key features: - World and character development - Arc planning and decision-making - RAG (Read-After-Generate) long-term memory - Chapter-by-chapter review and rollback - Production pipeline with checkpointing and rollback capabilities Tech stack: Go Why it matters: novel-studio offers a reliable, local-first approach to AI-generated fiction, ensuring the preservation of the story's integrity and avoiding the loss of context, making it a valuable tool for writers and creators looking to produce high-quality, verifiable long-form content. Repository: Xiaoyangy/novel-studio GitHub stars: 151 GitHub description: Open-source AI novel generator for long-form fiction. It provides a robust system for world-building, character development, and narrative progression, ensuring the preservation of the story's integrity and avoiding the loss of context, making it a valuable tool for writers and creators looking to produce high-quality, verifiable long-form content.

Overview: IRON is a Python library designed to simplify the programming of AMD Ryzen AI NPUs, enabling the execution of operators and models directly on the NPU, including Llama 3.2, Granite 4, Gemma 4, and EmbeddingGemma 2. It leverages MLIR-AIE for efficient compilation and autotuning. Key features: - Operators declared as Python classes with numpy-like APIs for easy integration. - Graphs of models compiled once per input shape, ensuring intermediate values remain on the device. - Autotuning of operators' tunables for optimal performance on each NPU variant. - Models run entirely on the NPU, with performance checks performed nightly by CI. Tech stack: MLIR-AIE bindings for AMD Ryzen AI NPUs, built on top of MLIR. Why it matters: IRON simplifies the development and deployment of AI models on AMD's NPUs, making it easier to integrate AI capabilities into existing systems without requiring deep knowledge of NPU-specific APIs or low-level programming.

### Overview: This project aims to fine-tune the Wav2vec 2.0 model for speech recognition tasks on datasets like Common Voice and Vivo. ### Key features: - Multi-GPU training - Automatic Mix Precision ### Tech stack: Python ### Why it matters: This project provides a simple and effective way to fine-tune the Wav2vec 2.0 model for speech recognition tasks, making it accessible and efficient for developers.

### Overview: Confucius is an official repository for the AAAI 2024 paper "Confucius: Iterative Tool Learning from Introspection Feedback by Easy-to-Difficult Curriculum". It aims to teach large language models (LLMs) how to use external tools, making them effective agents for practical task planning and tool calling. ### Key Features: - **Multi-stage Learning Method**: Teaches LLM to use various tools from an easy-to-difficult curriculum. - **Iterative Self-instruction from Introspective Feedback (ISIF)**: Dynamically constructs datasets to improve the ability to use complicated tools. - **Comprehensive Dataset**: Includes detailed prompts, tasks, and solutions for practical applications. ### Tech Stack: Python ### Why It Matters: This project addresses the need for LLMs to effectively integrate with external tools, enhancing their practical utility in real-world scenarios. The proposed method, Confucius, provides a novel approach to tool learning that addresses the limitations of existing methods by focusing on the complexity of tools in real-world applications.

Overview: This project is the official repository for the paper "How to build a consistency model: Learning flow maps via self-distillation" presented at NeurIPS 2025. It provides a unified mathematical framework for designing flow maps and introduces three direct training algorithms based on a notion of self-distillation. Key features: - Lagrangian Self-Distillation (LSD) - Progressive Self-Distillation (PSD) - Eulerian Self-Distillation (ESD) Tech stack: Jupyter Notebook Why it matters: This project offers a comprehensive solution for generating samples in generative modeling tasks with minimal forward passes, showcasing superior performance and training stability through the use of self-distillation methods.

Overview: TuneKit simplifies the fine-tuning process for LLMs, eliminating the need for local GPU setup, expensive hardware, and manual configuration. It provides a free, pre-configured Colab notebook that takes just 15 minutes to fine-tune a model. Key features: - Upload JSONL file - Answer two questions for optimal model and hyperparameters - Generate a pre-configured Colab notebook - Train on free T4 GPU in just 15 minutes - Download the fine-tuned model in various formats Tech stack: Python, FastAPI, Unsloth Why it matters: It eliminates the pain points of fine-tuning LLMs, making it accessible to developers without the need for expensive hardware or extensive coding. Repository: riyanshibohra/TuneKit Primary language: Python GitHub stars: 152 GitHub description: Upload your data → Get a fine-tuned SLM. Free.

Overview: Datatune is a powerful tool designed to transform and analyze large datasets using LLMs (Large Language Models) and Agents. It allows LLMs and Agents to access and manipulate infinite amounts of data, applying semantic intelligence to every record. Key features: - Data transformation and analysis - Integration with LLMs and Agents - Support for Dask and Ibis data sources - Row-level intelligence and semantic understanding Tech stack: Python Why it matters: Datatune simplifies the process of applying LLMs and Agents to large datasets, enabling them to access and manipulate data in real-time, improving efficiency and accuracy in data-driven decision-making. Repository: vitalops/datatune GitHub stars: 152 GitHub description: Agentic data transformation on infinite amounts of data GitHub license: MIT License GitHub repository: https://github.com/vitalops/datatune

### Overview: InstructRAG is a simple yet effective Retrieval-Augmented Generation (RAG) framework that leverages instruction-tuned language models to generate rationales for better verifiability and trustworthiness. This approach enhances the robustness and effectiveness of language models by explicitly denoising retrieved contents. ### Key Features: - **Self-Synthesis**: Generates its own supervision for denoising. - **Easy-to-Use**: Supports both in-context learning (ICL) and supervised fine-tuning (SFT). - **Effectiveness**: Up to 8.3% better results across 5 benchmarks. - **Noise Robustness**: Robust to increased noise ratios in various scenarios. - **Task Transferability**: Solves out-of-domain unseen tasks. ### Tech Stack: Python ### Why It Matters: InstructRAG significantly improves the reliability and accuracy of language models by providing explicit denoising mechanisms, making it a valuable tool for enhancing the performance of large language models in various applications.

### Overview: AIbitat is a TypeScript-based multi-agent conversation framework designed for enabling interaction between multiple agents while allowing human participation. ### Key Features: - Direct messages between agents - Feedback mechanism to keep the chat alive - Channels for agents to talk with many others - Error handling for rate limits - Function execution capability - Web browsing capability - File interaction capability - Code execution capability - Cost limit management - Providers: OpenAI, Anthropic, Cohere, Fireworks.ai, Hugging Face, Replicate ### Tech Stack: TypeScript ### Why it Matters: AIbitat simplifies the creation and management of multi-agent conversations, making it easier for developers to integrate AI into their projects while maintaining human participation in the interaction.

Overview: ppt-master-comfyui is a personal-machine fork of ppt-master (v6.7.0) that features a local ComfyUI image backend, enabling the default use of Qwen-Image-2.1 for AI slide imagery without requiring an API key or cost. It also offers selectable workflows for MOSS-Image SFT and legacy "zit基础" Z-Image Turbo. Key features: - Local ComfyUI image backend - Qwen-Image-2.1 default for AI slide imagery - MOSS-Image SFT for literal, layout-steerable art - Z-Image Turbo for speed Tech stack: Python Why it matters: This fork provides a local ComfyUI backend for generating editable PPTX files without any API key or cost, making it suitable for personal use and collaboration without restrictions.

### Overview: Voconly is a local-first AI voice typing tool that transcribes speech locally, refines it with advanced AI, and presents finished text directly in your editor. It supports Whisper, SenseVoice, Parakeet, and Qwen ASR, allowing you to work offline and enhance your writing with LLM post-processing. This tool eliminates the need for cloud storage, quotas, and copy-paste, making it perfect for productivity on the go. ### Key Features: - **Local Transcription:** Transcribes speech locally, ensuring privacy and no cloud storage. - **AI Refinement:** Refines the transcribed text with advanced AI, improving readability and quality. - **LLM Post-Processing:** Works seamlessly with major LLM providers, enhancing the final output. - **Two Modes:** Switch between "Light Polish" for quick notes and "Professional Polish" for polished emails, meetings, and translations. - **Hotkeys:** Assign hotkeys to different modes, allowing for flexible workflow customization. - **Cross-Platform:** Works across various applications like browsers, Word, Notion, VS Code, WeChat, and email clients. ### Tech Stack: - **C++:** The primary programming language used for the development of Voconly. ### Why It Matters: - **Privacy:** Ensures that audio data never leaves your device, providing complete privacy. - **No Quotas:** Offers unlimited usage with no cloud quotas or subscription fees. - **Accuracy:** Provides finished, polished text directly in your editor, eliminating the need for copy-paste. - **Long Sessions:**