Projects

### Overview: Whallm is a language model designed to run on Apple Silicon Macs, utilizing minimal RAM and fast SSDs for optimal performance. It supports models like Qwen3.8 and DeepSeek V4.1, allowing for efficient inference on any M-series MacBook. ### Key Features: - **Model Compatibility**: Supports Qwen3.8, DeepSeek V4.1, and other models. - **Memory Efficiency**: Reads common weights in memory and downloads experts from SSD as needed. - **Chat Interface**: Provides a user-friendly chat interface for direct communication. - **OpenAI API Compatibility**: Compatible with OpenAI APIs, including Codex, for seamless integration. ### Tech Stack: Whallm uses Python for its core functionality and relies on Apple Silicon for optimal performance. ### Why It Matters: Whallm is designed to run language models that require significant memory, making it ideal for users with limited RAM. By efficiently managing memory and leveraging SSDs, Whallm ensures smooth operation even on devices with limited RAM. ### Repository: Repository: yanun0323/Whallm Stars: 152 Description: DeepSeek-V4-Flash-0731 284B inference in ~25 GB of RAM / Qwen3.8-Next-Flash inference in ~15 GB of RAM on any M-series MacBook

### Overview: Mnemosyne is a local-first long-term memory system for AI agents. It uses a single SQLite file to store evidence and memories, ensuring that the agent starts from the most up-to-date and accurate information possible. Mnemosyne supports evidence retention and correction, making it an essential tool for maintaining a clean and effective knowledge base. ### Key Features: - **Evidence Retention and Correction:** Each memory keeps its source and trust tier, and superseded versions are still inspectable after corrections. - **Corrections that Reach Dependents:** Corrected requirements retire the procedures and skills that depend on it, ensuring the agent learns from the most recent and accurate information. - **Auditable Over Time:** Past beliefs replay exactly, allowing for precise tracking of changes and improvements. - **Any Host:** Supports MCP for Claude Code, Cursor, and Codex, as well as TypeScript SDK, CLI, and authenticated HTTP. ### Tech Stack: TypeScript ### Why It Matters: Mnemosyne ensures that AI agents start from the most accurate and evidence-backed information, preventing the repetition of stale lessons and maintaining a clean and effective knowledge base. Its local-first approach makes it a powerful tool for maintaining a robust and reliable AI system.

Overview: ceLLama is a streamlined automation pipeline for cell type annotations using large-language models (LLMs). It ensures privacy and speed with extensive customized reports. Key features: - Operates locally, ensuring no data leaks. - Considers negative genes. - Efficient processing. - Generates extensive, customized reports. Tech stack: R Why it matters: ceLLama is ideal for quick and preliminary cell type checks, ensuring privacy and speed with comprehensive analysis and extensive reporting. Repository: CelVoxes/ceLLama GitHub stars: 151

### Overview: GOAT: A novel framework that boosts Low-Rank Adaptation (LoRA) for large-scale model fine-tuning by integrating adaptive singular-value priors and theoretical optimization alignment within a Mixture-of-Experts (MoE) architecture. ### Key Features: - **Adaptive Priors Initialization**: Segments the pre-trained weight’s SVD into multiple chunks; initializes each MoE expert with a distinct singular-value segment, dynamically selected by the router. - **Theoretical Optimization Alignment**: Derives closed-form scaling factors and residual corrections to align LoRA’s low-rank adaptation gradients and weights with full fine-tuning trajectories. - **State-of-the-Art Performance**: Outperforms existing LoRA and LoRA-MoE baselines across 25 benchmarks in image classification, NLU, NLG, and commonsense reasoning, closing or exceeding full fine-tuning accuracy. ### Tech Stack: - Python ### Why it Matters: Our paper has been accepted by ICML2025, demonstrating the effectiveness and advancement of our framework in improving the performance of LoRA for large-scale model fine-tuning.

### Overview Intugle's GenAI-powered open-source Python library transforms fragmented data into a connected semantic data model, enabling automated data intelligence. This toolkit discovers meaningful links and relationships across data assets, enriching them with profiles, classifications, and business glossaries. The resulting connected knowledge layer simplifies data integration and exploration, making it faster, more accurate, and less manual. ### Key Features - **Semantic Data Model:** Transform raw, fragmented datasets into an intelligent semantic graph that captures entities, relationships, and context. - **Business Glossary & Semantic Search:** Auto-generate a business glossary and enable search that understands meaning, not just keywords. - **Data Products:** Instantly generate SQL and reusable data products enriched with context, eliminating manual pipelines. - **Conceptual Search:** Generate data product plans from natural language queries, bridging the gap between business questions and executable data product definitions. ### Tech Stack Jupyter Notebook ### Why It Matters Intugle simplifies data integration and exploration, making it faster, more accurate, and less manual. It transforms fragmented data into a connected semantic data model, enabling automated data intelligence. This toolkit discovers meaningful links and relationships across data assets, enriching them with profiles, classifications, and business glossaries. The resulting connected knowledge layer simplifies data integration and exploration, making it faster, more accurate, and less manual.

### Overview: This project involves speedrunning LoRA fine-tuning on Qwen2.5-1.5B, aiming to achieve a GSM8K accuracy of 57% or higher within a single L40S. The leaderboard features verified records with detailed analysis and techniques. ### Key features: - **Leaderboard**: Displays the current record and verified records with their train times, GSM8K/EM scores, and technique details. - **Track 1**: Uses GSM8K with Qwen2.5-1.5B, achieving a baseline of 59.4% with optimization. - **Track 2**: Ported to SQuAD v1.1 with SmolLM2-1.7B, achieving a baseline of 77.5% with optimization. ### Tech stack: Python ### Why it matters: This project showcases the performance of LoRA fine-tuning on large language models, providing benchmarks and techniques that can be used to optimize future fine-tuning efforts.

### Overview: Diff-Harmonization is a zero-shot image harmonization method that addresses the reliance on synthetic composite images in existing techniques. By utilizing a generative model prior, it enables the harmonization of images without the need for extensive training data, making it more versatile and adaptable to various scenarios. ### Key features: - **Generative Model Prior**: The method leverages a generative model to harmonize images, providing a more flexible and adaptable approach. - **Automatic Object Name and Imaging Conditions Generation**: By incorporating Google GenMini VLM, the prompts for object names and imaging conditions can be automatically generated, enhancing the method's applicability. - **Content Preservation**: The technique preserves the content structure and imaging descriptions, ensuring that the harmonized images maintain their original details and context. - **Performance Evaluation and Selection**: An automatic performance evaluation and result selection process is integrated, improving the reliability and quality of the harmonization results. ### Tech stack: Python ### Why it matters: Diff-Harmonization offers a more efficient and versatile solution for image harmonization, making it suitable for a wide range of applications without the need for extensive training data. This approach enhances the method's adaptability and generalization capabilities, making it a valuable contribution to the field of image processing and computer vision.

### Overview: Welcome to Awesome AI Models, your ultimate destination for the latest in AI models and large language models. This curated list highlights the cutting-edge advancements in generative AI and LLMs, perfect for developers looking to showcase their innovative projects. ### Key Features: - **AI Model Hosting Platforms**: Explore platforms like Hugging Face and GitHub Models for easy model building, sharing, and deployment. - **Top LLMs**: Discover the latest and greatest models from leading AI companies, including Qwen, 01-ai, Alibaba AI, and more. - **LLMs and Models**: Dive into the world of AI models, including models from Alibaba, Apple, Baichuan, BigCode, DataBricks, DeepSeek, and more. ### Tech Stack: The repository uses Markdown for documentation and is maintained by a community of contributors. ### Why it Matters: Stay ahead of the curve with the latest AI model innovations. Whether you're a developer looking to showcase your work or an enthusiast exploring the latest advancements, this list is your guide to cutting-edge AI technology.

### Overview: DrugAssist is a large language model designed for molecule optimization, utilizing a combination of natural language processing and molecular optimization techniques to enhance drug discovery and development processes. ### Key Features: - **Molecule Optimization**: Utilizes a large language model to guide the optimization of molecules. - **Dataset**: Available on Hugging Face for training and evaluation. - **Model**: Fine-tuned on the MolOpt-Instructions dataset using LoRA. - **Web Demo**: Provides a user-friendly interface for molecular optimization and evaluation. - **Quantized Model**: Available for deployment on devices with limited hardware resources. ### Tech Stack: - Python - Hugging Face ### Why It Matters: DrugAssist offers a novel approach to molecule optimization by leveraging the power of language models, potentially accelerating the drug discovery process and improving the efficiency of drug development. ### Repository: Repository: blazerye/DrugAssist GitHub Stars: 151 Primary Language: Python GitHub Description: [Briefings In Bioinformatics] DrugAssist: A Large Language Model for Molecule Optimization

### Overview: The `pd-bridge` project is a heterogeneous prefill/decode solution designed to accelerate inference on DeepSeek-V4-Flash models. It leverages two production inference engines: one on NVIDIA DGX Spark and another on Apple Silicon Mac Studio, over a plain 10GbE link. The project aims to bridge the gap between these two systems, enabling efficient inference without the need for expensive hardware or specialized software. ### Key Features: - **Prefill:** Prefills a 284B model on NVIDIA DGX Spark and vLLM TP2. - **Decode:** Decodes the same model on Apple Silicon Mac Studio using oMLX. - **Link:** Utilizes ordinary 10GbE for communication, no RDMA or Thunderbolt. ### Tech Stack: The project is built using Python, leveraging the `deepseek-ai/DeepSeek-V4-Flash` model prefill and decode libraries. ### Why It Matters: The `pd-bridge` project is crucial for accelerating inference on DeepSeek-V4-Flash models, particularly for applications that require high throughput and low latency. By bridging the gap between two different hardware platforms, it enables efficient inference without the need for expensive hardware or specialized software, making it a valuable tool for developers and researchers working with this type of model.

Overview: This repository contains a collection of small iOS projects developed with Swift 5.7, focusing on core concepts such as UIKit, navigation, persistence, networking, gestures, maps, and media. The projects are designed to reinforce Swift skills and align with modern iOS development standards. Key features: - Images - Segues - Gesture Recognizer - Colors - Launch Screen - Alert - Table View - Core Data - MapView - TouchID - Tab Bars - Intelligent Image - Navigation Bar - Toolbar - Kronometre - Klavyeyi Kontrol Etme - ViewController - Table View Example - Table View Search - Collection View - UIViewControllerLifecycle - TabBarApp - SwitchScreens - Delegate - NotificationCenter - FindMyAge - Calculator - Animations - Swipe And Shake App - Swipe Gesture Recognizer - Alert View - Picker View - DatePickerViewApp - Web View Tech stack: Swift 5.7 Why it matters: This repository serves as a practical resource for iOS developers, offering a collection of projects that reinforce core Swift and UIKit concepts. It is designed to be aligned with modern iOS development standards, ensuring that developers can learn and apply best practices in a practical setting.

Overview: NaRCan is a video editing framework that integrates a hybrid deformation field network with diffusion priors to maintain the natural appearance of the image while refining it for better editing. Key features: - Integrates a hybrid deformation field network - Uses diffusion priors for natural image generation - Supports video editing tasks Tech stack: Python Why it matters: NaRCan addresses the challenge of maintaining the natural appearance of images during video editing, making it suitable for applications requiring high-quality, natural-looking edits.

### Overview: Andromeda is an ultra-fast and intelligent language model designed to handle ultra-long sequences of up to 100,000+ context lengths. It is optimized for efficiency, flexibility, scalability, and community-driven development. ### Key Features: - **Ultra Long Sequences**: Supports sequences of up to 200,000+ context lengths. - **Ultra Fast Processing**: Processes sequences in under 100ms. - **Superior Reasoning Capabilities**: Provides enhanced reasoning capabilities. ### Tech Stack: - Python ### Why It Matters: Andromeda is designed to be the fastest, most creative, and reliable language model ever built, making it ideal for a wide range of applications. Its efficient processing and flexible adaptability make it a powerful tool for developers and researchers alike. ### Repository: kyegomez/Andromeda ### Primary Language: Python ### GitHub Stars: 151 ### Installation: ```bash python3.11 -m pip install --upgrade andromeda-torch ``` ### Usage: - **Forward pass with random inputs**: ```python import torch from andromeda.configs import Andromeda1Billion model = Andromeda1Billion() x = torch.randint(0, 256, (1, 1024)).cuda() out = model(x) # (1, 1024, 20000) print(out) ``` -

### Overview: Marqo-FashionCLIP and Marqo-FashionSigLIP are state-of-the-art CLIP/SigLIP embedding models fine-tuned specifically for the fashion domain. These models achieve a 57% improvement in evaluation metrics compared to FashionCLIP 2.0 across seven benchmark datasets. The repository includes benchmark results showing significant performance gains over pretrained OpenCLIP models and leading fashion CLIP models. ### Key Features: - Fine-tuned CLIP/SigLIP models for the fashion domain. - Averaged performance across seven benchmark datasets. - 57% improvement in evaluation metrics compared to FashionCLIP 2.0. - Models released on Hugging Face: Marqo-FashionCLIP and Marqo-FashionSigLIP. ### Tech Stack: Python ### Why It Matters: These models significantly enhance the accuracy and relevance of fashion search and classification, making them valuable tools for e-commerce and fashion-related applications.

Overview: MCPsurfr is a local web search server that reranks results locally, eliminating the need for API keys. It uses a small MobileBERT model to re-rank the best pages, fetching only the main text and clean, non-menu content. Key features: - No API keys: Searches use ddgs, a multiple engine search. - Ranked locally: Results are re-ranked locally using a small MobileBERT model. - Clean text: Only the main content is fetched, with browser impersonation. - Passages instead of pages: Returns only the paragraphs that answer the question. - Reads blocked sites: Fetches Stack Overflow, Wikipedia, GitHub, Hacker News, PyPI, npm, and crates.io. - Titles and dates: Every page and passage carries its title and publish date. Tech stack: Python Why it matters: MCPsurfr is a practical solution for developers who want to use their own machine learning models for web search without the hassle of API keys. It efficiently re-ranks results and fetches only the necessary content, making it ideal for local use cases.

Overview: `g4l` is a Python library that allows developers to run language models locally using the `llama.cpp` bindings. It provides a sister project to `@gpt4free`, offering additional features like text retrieval from documents. Key features: - Local language model inference - Support for various models, including `orca-mini-3b` and `Mistral-7B-Instruct-v0.2` - Integration with documents for enhanced context retrieval Tech stack: Python Why it matters: `g4l` streamlines the process of deploying AI models locally, making it accessible for developers without the need for internet connections or external providers, thus reducing latency and resource usage. GitHub stars: 151 Repository: xtekky/gpt4local

### Overview: DataMind is an open-source project that focuses on developing LLM-based data analysis agents. It aims to address the challenges of open-source language models in data analysis by providing empirical diagnosis, generalist agent scaling, process-level supervision, long-horizon evaluation, and unsupervised skill discovery. These features collectively contribute to creating more capable, scalable, and reliable data-analytic agents. ### Key Features: - Empirical Diagnosis - Generalist Agent Scaling - Process-Level Supervision - Long-Horizon Evaluation - Unsupervised Skill Discovery ### Tech Stack: Python ### Why It Matters: DataMind provides a systematic path toward creating more capable, scalable, and reliable data-analytic agents, addressing the limitations of open-source language models in data analysis.

### Overview: This repository contains a collection of must-read papers on large language models (LLMs) that focus on continual learning. The papers explore various aspects of continual learning, including knowledge transfer, parameter-efficient tuning, and generalization to new data. ### Key Features: - **Towards Continual Knowledge Learning of Language Models** *Joel Jang, Seonghyeon Ye, Sohee Yang, Joongbo Shin, Janghoon Han, Gyeonghun Kim, Stanley Jungkyu Choi, Minjoon Seo* [[abs](https://arxiv.org/abs/2110.03215)]. ICLR 2022. - **Continual Pre-Training Mitigates Forgetting in Language and Vision** *Andrea Cossu, Tinne Tuytelaars, Antonio Carta, Lucia Passaro, Vincenzo Lomonaco, Davide Bacci.* [[abs](https://arxiv.org/abs/2205.09357)]. Preprint 2022.05. - **Lifelong Pretraining: Continually Adapting Language Models to Emerging Corpora** *Xisen Jin, Dejiao Zhang, Henghui Zhu, Wei Xiao, Shang-Wen Li, Xiaokai Wei, Andrew Arnold, Xiang Ren.* [[abs](https://arxiv.org/abs/2110.08534)]. NAACL 2022. - **Contin

Overview: iloveAgents is a community-driven, open-source web platform that allows users to run AI agents directly in their browsers. It supports multiple AI providers and offers features like Battle Mode, Workflow Builder, and Scheduled Automations. Key features: - Individual Agents - Battle Mode - Workflow Builder - Scheduled Automations Why it matters: iloveAgents provides a clean, easy-to-use interface for running AI agents directly in the browser, without the need for external servers. The platform's modular architecture makes it simple to add new agents, and the community-driven approach ensures that it remains a valuable resource for developers. Tech stack: JavaScript Why it matters: The use of JavaScript as the primary language allows for a clean, straightforward implementation of the platform's features. This makes it accessible to a wide range of developers and users. GitHub stars: 150 Why it matters: The high number of stars indicates that the project has gained significant traction and interest from the community, suggesting that it is meeting the needs of its users effectively.

Overview: InstructERC is an open-source implementation of a novel approach to reformulate the ERC task from a discriminative framework to a generative framework using large language models (LLMs). This project introduces a simple yet effective retrieval template module and two additional emotion alignment tasks to improve the performance on commonly used ERC datasets. Key features: - InstructERC's innovative multi-granularity dialogue supervision information integration. - Introduction of two emotion alignment tasks: speaker identification and emotion prediction. - Significantly outperforms previous models on three ERC datasets. Tech stack: Python Why it matters: InstructERC offers a practical solution for improving emotion recognition in conversational AI, making it valuable for developers working on conversational AI applications.