Today's Open Source (2024-07-29): Mistral Lightweight Model for Mobile Devices
Discover cutting-edge AI models: Lite-Oute-1 series, Llama3-Chinese, Agent-E, GraphRAG Local Ollama, Album AI, and Lyzr Automata. Enhance your projects today!
Here are some interesting AI open-source models and frameworks I wanted to share today:
Project: Lite-Oute-1
Lite-Oute-1 is a lightweight open-source model series with 65M, 150M, and 300M parameters.
The 65M model, based on the Llama architecture, has basic text generation capabilities but needs improvement in instruction execution and topic coherence.
The 150M model uses the Mistral architecture and can run on mobile devices, but may struggle with context coherence in multi-turn conversations.
The 300M model builds on the 150M, with more parameters and finer training data, aiming to enhance performance while maintaining cross-device efficiency.
https://huggingface.co/OuteAI/Lite-Oute-1-300M
https://huggingface.co/OuteAI/Lite-Mistral-150M-v2-Instruct
https://huggingface.co/OuteAI/Lite-Oute-1-65M
Project: Llama3-Chinese
The Llama3.1 Chinese fine-tuned model is an open-source version based on Llama3.1-8b-instruct with DPO fine-tuning. It can answer questions in fun Chinese with emojis, similar to the original instructions.
https://huggingface.co/shareAI/llama3.1-8b-instruct-dpo-zh
https://huggingface.co/shareAI/llama3-Chinese-chat-8b
https://huggingface.co/shareAI/llama3-8b-Chinese-Instruct-DPO-beta0.5
Project: Agent-E
Agent-E is an open-source Agent framework for automating operations on users' computers.
It focuses on browser automation and uses the AutoGen agent framework for natural language interaction with web browsers.
It can fill forms, search and sort products on e-commerce sites, locate specific content, manage media playback, perform web searches, and manage tasks on project management platforms.
https://arxiv.org/abs/2407.13032
https://github.com/EmergenceAI/Agent-E
Project: GraphRAG Local Ollama
GraphRAG Local Ollama is a modified version of Microsoft's GraphRAG, supporting local models downloaded from Ollama.
The project aims to replace expensive OpenAPI models with efficient local inference.
GraphRAG combines Retrieval-Augmented Generation (RAG) and graphical text indexing to answer global questions about private text corpora.
https://github.com/TheAiSingularity/graphrag-local-ollama
Project: Album AI
Album AI is an experimental project that automatically identifies the metadata of images in albums and uses RAG technology for album interaction.
It uses the latest gpt-4o-mini and Haiku as visual models, functioning as a traditional photo album or image knowledge base to help large language models generate content.
https://github.com/gcui-art/album-ai
Project: Lyzr Automata
Lyzr Automata is a low-code multi-agent automation framework designed to improve workflow efficiency and effectiveness.
It allows the creation of multiple Agents coupled with specific tasks, which can run independently to complete given instructions and achieve stability.
The framework supports connecting different large models and provides pre-built model classes and tools. Users can also extend base classes to support custom models and tools.
