Today's Open Source (2024-08-20): Llama-3.1-Storm-8B Boosts Instruction Following with Self-Curated Data
Today's Open Source (2024-08-20): Llama-3.1-Storm-8B Boosts Instruction Following with Self-Curated Data
Discover the latest in AI open-source models, including Llama-3.1-Storm-8B and Magnum v2 123b, designed to enhance language tasks and intelligent search.
Here are some interesting AI open-source models and frameworks I wanted to share today:
Project: Llama-3.1-Storm-8B
Llama-3.1-Storm-8B is an advanced language model designed to enhance the performance of 8B parameter models across various tasks.
Built on Llama-3.1-8B-Instruct, it significantly improves instruction following and function calling through techniques like self-curated high-quality data, targeted fine-tuning, and model fusion.
The high-quality data was curated from the NeurIPS 2023 challenge winners. About 975,000 training examples were selected based on educational value and difficulty from vast open-source datasets. The model itself evaluated the data quality, increasing the relevance and effectiveness of the training set.
https://huggingface.co/blog/akjindal53244/llama31-storm8b
https://huggingface.co/akjindal53244/Llama-3.1-Storm-8B
Project: Magnum v2 123b
Magnum-v2-123b is a model designed to replicate the prose quality of Claude 3, specifically the Sonnet and Opus styles.
It is fine-tuned on the Mistral-Large-Instruct-2407 base, supports up to 9 languages, and is suitable for text generation and conversation tasks.
https://huggingface.co/anthracite-org/magnum-v2-123b
https://huggingface.co/anthracite-org/magnum-v2-123b-gguf
https://huggingface.co/anthracite-org/magnum-v2-123b-exl2
Project: Controllable-RAG-Agent
Controllable-RAG-Agent is an advanced retrieval-augmented generation (RAG) solution designed to tackle complex problems that simple semantic similarity retrieval cannot solve.
This project features a detailed deterministic graph as the "mind" of an autonomous agent, capable of answering non-trivial questions from custom datasets. It includes features like hallucination prevention, multi-step reasoning, and adaptive planning.
https://github.com/NirDiamant/Controllable-RAG-Agent
Project: RAGMeUp
RAG Me Up is a versatile framework (including server and UI) that allows users to easily perform RAG (retrieval-augmented generation) on their own datasets.
At its core is a small, lightweight server and several methods to run the user interface and communicate with the server. RAG Me Up can run on a CPU, but for optimal performance with the default instruction model, it's recommended to use a GPU with at least 16GB of vRAM.
https://github.com/AI-Commandos/RAGMeUp
Project: Sensei
Sensei is an open-source AI-driven answer engine similar to Perplexity, using multiple open-source large language models (LLMs) for smart search and answer capabilities.
This project uses a modern tech stack for both front-end and back-end, supporting deployment both locally and in the cloud. It is ideal for applications needing efficient search and question-answering features.
https://github.com/jjleng/sensei
Project: ADAS
ADAS (Automated Design Agent System) is a research project aimed at automatically creating robust agent system designs. It explores new building blocks for inventing and combining these systems.
At the core of ADAS is a simple yet effective algorithm called Meta Agent Search. This algorithm allows a "meta" agent to iteratively program new agents based on previous discoveries, promoting the invention of novel and powerful agent designs.
