Projects

### Overview: Xolotl is a local coding agent optimized for one gaming PC, featuring the DeepSeek-V4.1 engine, GLM-5.3-Flash, and Qwen3.8 models. It achieves up to 27 tok/s on coding tasks, with a context size of 512K, and supports up to 90 GB of RAM and 90 GB of NVMe storage. ### Key Features: - **DeepSeek-V4.1 Engine**: The core engine with 552 billion parameters. - **GLM-5.3-Flash Engine**: Enhances the model with GLM-5.3-Flash. - **Qwen3.8 Models**: Includes Qwen3.8-27B and Qwen3.8-Next, with expert-tiering. - **OpenAI and Anthropic API Compatibility**: Compatible with OpenAI and Anthropic APIs. ### Tech Stack: - **C++**: The primary programming language used. - **Linux**: The operating system. - **NVIDIA**: GPU support. ### Why It Matters: Xolotl is a significant improvement over the original Strata-DSV4.1, offering enhanced performance and compatibility with modern APIs. It's optimized for gaming PCs, making it ideal for developers and coders who need a powerful yet lightweight solution. ### GitHub Stars: 380 ### Repository: Indras-Mirror/Xolotl ### Primary Language: C++ ### GitHub Description: Xolotl (formerly Strata-DS

### Overview: BERT-NER (nert-bert) is a project that utilizes the pre-trained BERT model from Google to perform Named Entity Recognition (NER) tasks. This implementation is designed to be easily adaptable and can be fine-tuned with any pre-trained TensorFlow checkpoint. ### Key features: - **BERT-NER Model**: A pre-trained BERT model is used for NER tasks. - **Fine-tuning with TensorFlow checkpoints**: The model can be fine-tuned with any pre-trained TensorFlow checkpoint, making it highly adaptable. - **Data Preparation**: The project includes a method to create data for training and evaluation, including creating datasets and data loaders. - **Model Creation**: The project provides a method to create models using the BERT-NER architecture. - **Prediction and Evaluation**: The project includes methods for predicting and evaluating the model's performance, including token and span metrics. ### Tech stack: The project uses Python with Jupyter Notebook for implementation and data preparation. ### Why it matters: BERT-NER is a significant contribution to the NER field, offering a pre-trained BERT model that can be fine-tuned for specific tasks, making it highly adaptable and efficient for various NER applications.

Overview: Rosie is a Python application designed to enhance the intelligence of Serenata de Amor, a project focused on promoting love and unity. Key features: - Integrates with Serenata de Amor's existing infrastructure. - Enhances data analysis and machine learning capabilities. - Improves user engagement and interaction with the platform. Tech stack: Python Why it matters: By enhancing the intelligence of Serenata de Amor, Rosie ensures a more informed and engaging user experience, ultimately fostering a deeper connection and appreciation for love and unity.

### Overview: Veldra is a self-hostable, local-first agent platform that allows developers to create and manage agents. It compiles natural language descriptions into working agents, including policies, tools, reasoning methods, and teams. Agents learn from feedback and improve over time, making them adaptable and efficient. ### Key Features: - **Smart Composer**: Routes tasks to the best existing agent or builds new ones based on prompts. - **Real Workspace**: A first-run install wizard for workspace setup, and a sign-in gate for teammates. - **MCP Connectors**: Built-in tools and external connectors (e.g., Shopify, Alibaba) that can be configured and enabled with explicit approval. - **Agent Teams**: Orchestrates teams with roles (admin, member, viewer) for seamless collaboration. - **Editable Knowledge Bases**: Supports uploading PDFs, indexing web pages, and editing documents with re-embedding capabilities. - **Visual Workflow Builder**: A visual builder for creating workflows, including conditional logic, code sandbox, and HTTP templates. - **Agents with a Soul**: Agents have personas that shape their interactions, and they grow and improve based on feedback and mistakes. - **Faust**: A floating admin bot for renaming, tagging, and managing agents, with audit logs and clear deletion options. ### Tech Stack: Python ### Why It Matters: Veldra simplifies the process of creating and managing agents, making it accessible and efficient for developers. Its ability to learn from feedback and adapt over time makes it a valuable tool for automating tasks and improving processes.

### Overview InfLLM is a training-free memory-based method designed to enhance the processing of extremely long sequences by large language models (LLMs). Unlike traditional approaches that discard distant tokens, InfLLM stores distant contexts in additional memory units and efficiently computes attention using an efficient mechanism. This allows LLMs pre-trained on sequences of a few thousand tokens to achieve superior performance compared to competitive baselines that continuously train on longer sequences. InfLLM maintains the ability to capture long-distance dependencies, making it a powerful tool for understanding and processing long sequences. ### Key Features - **Training-Free Memory**: InfLLM stores distant contexts in additional memory units and efficiently computes attention using an efficient mechanism. - **Efficient Attention Computation**: Utilizes an efficient mechanism to lookup token-relevant units for attention computation, enabling LLMs to process long sequences. - **Intrinsic Ability**: Allows LLMs pre-trained on sequences of a few thousand tokens to achieve superior performance compared to competitive baselines. - **Supports Topk Retrieval**: InfLLM supports topk retrieval using Faiss, improving inference speed and reducing GPU memory usage. ### Tech Stack - **Python**: The primary language used for developing InfLLM. - **GitHub Stars**: 407 - **GitHub Description**: The code of our paper "InfLLM: Unveiling the Intrinsic Capacity of LLMs for Understanding Extremely Long Sequences with Training-Free Memory". ### Why It Matters InfLLM provides a significant enhancement in the processing of extremely long sequences by LLMs, enabling

### Overview: Zeraix is an open-source local AI workspace designed for on-device inference, enabling users to run supported models locally, work with files and tools, and delegate tasks to specialized sub-agents. The project is built on a Rust Agent Runtime and is available in multiple operating systems. ### Key Features: - **Local Models**: Download and run supported GGUF models with hardware-aware recommendations and runtime management. - **File Operations**: Work with files and tools within the same workspace, including file operations and shared formatting guidance. - **Agents**: Delegate work to specialized sub-agents, inspect their execution, and control them with different modes. - **Local Workflows**: Keep local conversations and memory, track task goals, and schedule recurring workflows with run history and notifications. - **Document Tools**: Use Skills for Word documents, PDFs, presentations, and spreadsheets with shared formatting guidance. - **Media Library**: Collect uploaded and generated assets in a media library, preview images, video, audio, PDFs, and text, and connect compatible image or video generation services. - **Plugin Catalogue**: Connect local or remote MCP servers, browse the plugin catalogue, or bring your own model endpoint. - **QEMU Sandbox**: An optional isolated environment for supported commands. ### Tech Stack: Rust Agent Runtime ### Why It Matters: - **Local Models**: Run models locally without the need for a Zeraix account, subscription, or API key. - **File Operations**: Seamless integration with file operations and shared formatting guidance. - **Agents**: Delegate tasks to specialized sub-agents with

### Overview: EasyInstruct is an easy-to-use framework designed for Large Language Models (LLMs) like GPT-4, LLaMA, and ChatGLM. It simplifies instruction generation, selection, and prompting, making it accessible for research experiments. ### Key Features: - **Instruction Generation**: Easy-to-use methods for generating instructions. - **Instruction Selection**: Tools for selecting the best instructions. - **Prompting**: Framework for creating prompts for LLMs. - **Modularization**: Separates instruction generation, selection, and prompting for easy integration. ### Tech Stack: - Python ### Why It Matters: EasyInstruct streamlines the process of using LLMs in research, making it more accessible and efficient for developers and researchers. ### Repository: - Repository: zjunlp/EasyInstruct - GitHub Stars: 407 - GitHub Description: [ACL 2024] An Easy-to-use Instruction Processing Framework for LLMs. ### Table of Contents: - **What's New** - **Overview** - **Installation** - **Quickstart** - **Shell Script** - **Gradio App** - **Use EasyInstruct** - **Generators** - **Selectors** - **Prompts** - **Engines** - **Citation** - **Contributors** ### 🔔News: - **2024-06-04**: Accepted by ACL 2024 System Demonstration Track. - **2024-02-0

Overview: This TypeScript SDK enables developers to easily build AI agents that seamlessly integrate with multiple MCP tools, including OpenAI, Claude, Mistral, and others. It supports automatic tool selection, multi-agent crews, conversational results, advanced patterns, streaming, type safety, observability, and production-ready features. Key features: - Automatic tool selection - Multi-agent crews - Conversational results - 100s of models - Advanced patterns - Streaming - TypeScript-first approach - Observability - Production-ready Tech stack: TypeScript Why it matters: This SDK simplifies the integration of AI agents with real-world actions, making it easier for developers to create complex workflows without the need for manual configuration or extensive setup. Repository: Kong/volcano-agent-sdk GitHub stars: 406

Overview: This repository contains a comprehensive guide for rebutting papers at major AI conferences, focusing on common scenarios and tips for effective rebuttals. Key features: - Tips for addressing innovation, motivation, and theoretical boundaries - Tips for clarifying writing, related work, and communication strategies - Tips for experimental evidence and benchmarking Tech stack: Python, Markdown, and HTML Why it matters: This guide helps researchers and reviewers effectively respond to rebuttals, ensuring that their work is well-received and understood by the community. Repository: MLNLP-World/Paper-Rebuttal-Tips GitHub stars: 398

### Overview: Unlocking deep, interpretable biological reasoning from complex genomic data is a major AI challenge hindering scientific discovery. Current DNA foundation models struggle with multi-step reasoning and lack inherent transparent, biologically intuitive explanations. BIOREASON, a pioneering architecture, integrates a DNA foundation model with a large language model (LLM), enabling the LLM to directly process and reason with genomic information as a fundamental input, fostering a new form of multimodal biological understanding. BIOREASON's sophisticated multi-step reasoning is developed through supervised fine-tuning and targeted reinforcement learning, guiding the system to generate logical, biologically coherent deductions. Across biological reasoning benchmarks, BIOREASON significantly improves performance, raising accuracy on KEGG-based disease pathway prediction from 86% to 98% and delivering an average 15% gain over strong single-modality baselines in variant effect prediction tasks. BIOREASON reasons over unseen biological entities and articulates decision-making through interpretable, step-by-step biological traces, offering a transformative approach for AI in biology that enables deeper mechanistic insights and accelerates testable hypothesis generation from genomic data. ### Key Features: - **DNA Foundation Model Integration**: BIOREASON seamlessly integrates a DNA foundation model with a large language model (LLM), enabling the LLM to process and reason with genomic information. - **Multi-Step Reasoning**: BIOREASON's reasoning methodology is developed through supervised fine-tuning and targeted reinforcement learning, guiding the system to generate logical, biologically coherent deductions. - **Benchmark Development**: The architecture has been developed and curation of

### Overview: Q-Transformer is an implementation of a scalable offline reinforcement learning method that utilizes autoregressive Q-functions. This approach aims to address the scalability issues inherent in traditional Q-learning by leveraging autoregressive models to predict future states based on past actions. The project is part of the ongoing research by Google DeepMind and is designed to be a comprehensive guide for developers interested in understanding and implementing Q-learning techniques. ### Key Features: - **Autoregressive Q-Functions**: The project focuses on autoregressive Q-learning, which is a variant of Q-learning that predicts the optimal action based on the sequence of actions taken so far. - **Scalability**: The implementation is designed to be scalable, making it suitable for complex environments where traditional Q-learning methods may struggle. - **Single Action Support**: The project includes logic for Q-learning on single actions, which is essential for final comparisons with autoregressive Q-learning. - **Education**: The project aims to provide educational value by serving as a reference for implementing Q-learning on single actions and by comparing it with autoregressive Q-learning. ### Tech Stack: - **Programming Language**: Python - **GitHub Stars**: 406 ### Why It Matters: - **Scalability in Reinforcement Learning**: The autoregressive Q-learning formulation addresses the scalability issues of traditional Q-learning by predicting future states based on past actions, making it suitable for complex environments. - **Educational Value**: The project serves as a valuable resource for developers looking to implement Q-learning techniques, providing a clear and comprehensive guide to the implementation. ###

### Overview: Zshot is a Python framework designed for zero-shot and few-shot named entity recognition and relationship recognition. It leverages language descriptions to generalize to unseen domains and is highly customizable for various applications such as mentions extraction, wikification, and zero-shot NER and RE extraction. ### Key Features: - **Zero and Few Shot Named Entity Recognition (NER):** Generalize to unseen domains using language descriptions. - **Zero and Few Shot Named Relationship Recognition:** Recognize relationships in unseen domains. - **Visualization:** Extract mentions and relationships for zero-shot NER and RE extraction. - **Customizability:** Highly customizable for various applications. ### Tech Stack: Python ### Why It Matters: Zshot is a powerful tool for researchers and developers looking to perform zero-shot and few-shot named entity recognition and relationship recognition. Its ability to generalize to unseen domains makes it particularly useful in applications where domain-specific knowledge is limited.

### Overview: AURA is a production-tested SRE agent platform designed for rapid deployment, offering comprehensive guardrails, APIs, state management, streaming, and failure handling to ensure AI models operate safely in production infrastructure. ### Key Features: - **Rust Implementation**: Built with Rust for high performance and reliability. - **Quick Start**: Easy installation and setup with a single script. - **Agent Initialization**: Initializes the agent with a specified language model provider and initial configuration. - **Customization**: Allows customization of existing agents or addition of new ones to fit specific infrastructure and SRE workflows. - **Security Boundary**: Operates within secure boundaries, allowing for air-gapped and highly regulated environments. - **Model Provider Flexibility**: Supports multiple model providers like OpenAI, Anthropic, Bedrock, Gemini, Ollama, and OpenRouter. - **Human Approval**: Requires human approval for sensitive actions before tool calls execute. - **Traceability**: Generates OpenTelemetry traces for detailed investigation of model, tool, and orchestration decisions. ### Tech Stack: Rust ### Why It Matters: AURA simplifies the integration of AI models into production environments, ensuring that AI operations are secure, reliable, and compliant with security boundaries. Its customizable nature and support for various model providers make it a versatile tool for SRE teams.

Overview: This repository contains end-to-end recipes for pre-training and fine-tuning BERT using Azure Machine Learning Service. Key features: - Pre-training BERT on a corpus - Fine-tuning an existing BERT model on a specialized task - Preprocessing scripts for domain-specific data Tech stack: Jupyter Notebook Why it matters: This repository provides a stable, predictable workflow for researchers to explore and experiment with building their own language representation models beyond BERT, using GPUs or Azure Machine Learning service.

### Overview: Quant.cpp is a pure C library designed for efficient LLM inference, providing a 7x longer context window compared to traditional RAG systems. It offers lossless KV cache compression and a single-header library, making it perfect for consumer hardware. ### Key Features: - **Lossless KV Cache Compression**: Ensures full-document understanding with 6.4x compression. - **Pure C Library**: No dependencies, making it lightweight and easy to integrate. - **128K Context**: Supports 128 FP32 tokens and 4-bit everything else, achieving FP32 quality. ### Tech Stack: - **C**: Pure C library with no external dependencies. ### Why It Matters: - **Efficient Inference**: Provides 7x longer context, enabling better understanding and faster inference times. - **Lightweight**: No dependencies, making it ideal for small devices and embedded systems. - **High Performance**: Supports 128 FP32 tokens, achieving FP32 quality at 3x compression, making it suitable for high-performance computing. ### Repository: - **Repository**: quantumaikr/quant.cpp - **GitHub Stars**: 405 - **Primary Language**: C ### Quick Start: - **CLI (v0.12.0+)**: ```bash pip install quantcpp quantcpp pull qwen3 # download Qwen3-4B Q4_K_M (~2.5 GB) quantcpp run qwen3 # interactive chat quantcpp

### Overview: This repository contains a comprehensive list of free resources for learning data science, covering Python, data science general, machine learning, natural language processing, deep learning, and mathematics/statistics. ### Key Features: - Python courses and tutorials from DataCamp, Dataquest, and CS50x. - Books on Python, machine learning, natural language processing, and deep learning. - Tutorials and resources for data science general, machine learning, natural language processing, deep learning, and mathematics/statistics. ### Tech Stack: Python, Data Science, Machine Learning, Natural Language Processing, Deep Learning, Mathematics/Statistics. ### Why it Matters: These resources provide a wealth of information and practical experience for aspiring data scientists, covering the foundational skills and advanced techniques needed in the field.

### Overview: LiBai is a comprehensive large-scale model training toolbox based on OneFlow, designed to facilitate efficient and effective parallel training of models across various domains including Computer Vision (CV) and Natural Language Processing (NLP). ### Key Features: - **Parallel Training Components**: LiBai supports Data Parallelism, Tensor Parallelism, and Pipeline Parallelism, making it extensible for new parallel training techniques. - **Training Techniques**: It includes out-of-the-box training methods such as Distributed Training, Mixed Precision Training, Activation Checkpointing, Recomputation, Gradient Accumulation, and Zero Redundancy Optimizer (ZeRO). - **Versatile Support**: LiBai is designed to work with both Computer Vision and Natural Language Processing datasets, providing pre-defined data processes for CIFAR, ImageNet, and BERT datasets. - **Ease of Use**: The components are modular, allowing for easier integration and usage. It is designed as a library to support building research projects, with features like LazyConfig for flexible syntax and no predefined structures, a friendly trainer and engine, and the ability to be used as a library. - **High Efficiency**: LiBai aims to provide high efficiency in training processes, optimizing user experience and performance. ### Tech Stack: Python ### Why it Matters: LiBai is essential for researchers and developers looking to efficiently train large-scale models, making it easier to scale training processes and accelerate research and development in fields such as Computer Vision and Natural Language Processing.

Overview: NLPnet is a Python library designed for performing Natural Language Processing tasks such as part-of-speech tagging, semantic role labeling, and dependency parsing using neural networks. It leverages Cython for enhanced performance and supports language-independent functionality with some optimizations for Portuguese. Key features: - Neural network-based POS tagging - Semantic role labeling - Dependency parsing Tech stack: Python with Cython for performance optimization. Why it matters: NLPnet provides a robust solution for NLP tasks, particularly useful for applications requiring fast and efficient processing of natural language data. Its language-independent architecture makes it versatile for various languages, including Portuguese, where it excels due to tailored optimizations. Repository: erickrf/nlpnet GitHub stars: 404 GitHub description: A neural network architecture for NLP tasks, using Cython for fast performance. Currently, it can perform POS tagging, SRL and dependency parsing. Gitter is chat room for developers.

Overview: AgentPrism is an open-source library of React components designed to visualize traces from AI agents, providing a clear and visual way to debug AI agents using traces. Key features: - Trace visualization - Hierarchical timeline - Integration with OpenTelemetry data - Support for LLM calls, tool executions, and agent workflows Tech stack: TypeScript Why it matters: AgentPrism simplifies the process of debugging AI agents by converting complex traces into easily understandable visual diagrams, making it a valuable tool for developers and researchers working with AI and machine learning systems. Repository: evilmartians/agent-prism Primary language: TypeScript GitHub stars: 404 GitHub description: React components for visualizing traces from AI agents README excerpt:

Overview: FreshLLMs is a project that provides data and code for the paper "FreshLLMs: Refreshing Large Language Models with Search Engine Augmentation," which was published on arXiv. The project includes a dataset called FreshQA, which is updated weekly or upon request, and a new evaluation benchmark called SealQA, which builds on FreshQA with a stronger focus on reasoning. Key features: - Data and code for the paper "FreshLLMs: Refreshing Large Language Models with Search Engine Augmentation" (arXiv:2310.03214). - Dataset called FreshQA, which is updated weekly or upon request. - New evaluation benchmark called SealQA, which builds on FreshQA with a stronger focus on reasoning. - Repository: freshllms/freshqa, primary language: Jupyter Notebook, GitHub stars: 404. Tech stack: Jupyter Notebook. Why it matters: FreshLLMs and SealQA provide valuable resources for researchers and developers interested in large language models, specifically focusing on the refreshing of models with search engine augmentation and the development of a new evaluation benchmark for reasoning.