Today's Open Source (2024-08-23): Jamba 1.5, 256K Context, Function Calls, JSON Output
Discover top AI open-source projects like Jamba 1.5, OpenHands, LongVILA, and more. Enhance development efficiency with cutting-edge tools and frameworks.
Here are some interesting AI open-source models and frameworks I wanted to share today:
Project: Jamba 1.5
The AI21 Jamba 1.5 series models are advanced hybrid SSM-Transformer instruction-following models. They support 256K context and offer 2.5 times faster inference than similar leading models.
Jamba 1.5 Mini (12B active/52B total parameters) and Jamba 1.5 Large (94B active/398B total parameters) are optimized for commercial use. They support functions like function calls, structured outputs (JSON), and basic text generation.
The models are released under the Jamba Open Model License, allowing commercial use.
https://huggingface.co/ai21labs/AI21-Jamba-1.5-Large
https://huggingface.co/ai21labs/AI21-Jamba-1.5-Mini
Project: OpenHands
OpenHands is a self-driven software engineering platform powered by large language models (LLMs).
OpenHands agents collaborate with developers to write code, fix bugs, and release features. The platform aims to boost development efficiency by reducing manual coding tasks.
https://github.com/All-Hands-AI/OpenHands
Project: LongVILA
LongVILA is a long-context visual language solution designed for multimodal foundation models.
It focuses on three areas: systems, model training, and dataset development.
For model training, LongVILA introduces a five-stage pipeline: alignment, pre-training, short-term supervised fine-tuning, context expansion, and long-term supervised fine-tuning.
This full-stack solution increases VILA's processing frames by 128 times, from 8 to 1024 frames. It also improves long video captioning scores from 2.00 to 3.26, achieving 99.5% accuracy in the "needle-in-a-haystack" test at 1400 frames (context length 274k).
https://github.com/NVlabs/VILA/blob/main/LongVILA.md
https://arxiv.org/abs/2408.10188
Project: HumanLayer
HumanLayer is a Python toolkit that allows AI Agents to communicate with humans in tool-based and asynchronous workflows.
It sets up approval workflows on platforms like Slack and email to ensure human oversight of high-risk function calls.
Users can choose their preferred LLM and framework to give AI Agents safe access to the world.
https://github.com/humanlayer/humanlayer
Project: CyberScraper 2077
CyberScraper 2077 is a powerful web scraping tool that uses advanced models like GPT-4 to intelligently parse and extract web content.
It offers a user-friendly Streamlit interface, supports multiple export formats (e.g., JSON, CSV, HTML, SQL, Excel), and has a stealth mode to avoid bot detection.
The project also supports asynchronous operations and smart parsing, with future plans to add Agent mode and multi-page scraping.
https://github.com/itsOwen/CyberScraper-2077
Project: StructuredRAG
StructuredRAG Benchmarker is a tool for evaluating LLMs' ability to generate output in specific JSON template formats.
The project compares f-String prompts with the Follow the Format (FF) method used in DSPy and explores structured output in various RAG tasks.
