Today's Open Source (2024-08-14): AI Scientist by Transformer Co-Creator
Explore today's top AI open-source releases: InternLM2.5-1.8B, FalconMamba 7B, AI Scientist, LLM-Aided OCR, RAGoon, and AgileCoder.
Here are some interesting AI open-source models and frameworks I wanted to share today:
Project: InternLM2.5/InternLM2.5-1.8B
InternLM2.5 series has released the open-source InternLM2.5 1.8B, with 32K context support. It significantly improves inference performance compared to InternLM2 1.8B.
InternLM 2.5 is a model series from Shanghai AI Lab, which previously released 7B/20B versions. It excels in instruction understanding, tool selection, and result reflection, gathering and analyzing information from hundreds of web pages.
https://github.com/InternLM/InternLM
https://huggingface.co/internlm/internlm2_5-7b
https://huggingface.co/internlm/internlm2_5-20b
Project: FalconMamba 7B
Falcon-Mamba-7B, developed by TII, is an autoregressive decoder model mainly for English text generation. Using the Mamba architecture and a multi-stage training strategy, it handles 8K context.
Trained on high-quality technical, code, and math data, it aims to deliver efficient and high-quality text generation.
https://huggingface.co/tiiuae/falcon-mamba-7b
https://huggingface.co/tiiuae/falcon-mamba-7b-instruct
Project: AI Scientist
The AI Scientist is a system for fully automated scientific discovery, developed by Sakana AI, a startup co-founded by Llion Jones, one of the creators of Transformer. It applies LLMs to independent research.
In demos, AI Scientists explored various machine learning subfields and made new contributions to popular areas like diffusion models, Transformers, and grokking.
https://github.com/SakanaAI/AI-Scientist
https://arxiv.org/abs/2408.06292
Project: LLM-Aided OCR
LLM-Aided OCR is an open-source project that enhances OCR output quality using LLMs.
The main idea is to use LLMs to correct and refine the raw OCR output, improving the text for spelling, grammar, and readability.
This project transforms raw OCR text into highly accurate, well-formatted, and readable documents.
https://github.com/Dicklesworthstone/llm_aided_ocr
Project: RAGoon
RAGoon is an advanced library for batch embedding generation, supporting fast web-based RAG (retrieval-augmented generation) and quantized indexing.
It offers NLP tools for multi-model embedding generation, and high-dimensional vector visualization, and enhances language model performance through query search, web scraping, and data augmentation.
https://github.com/louisbrulenaudet/ragoon
Project: AgileCoder
AgileCoder is an innovative multi-agent framework for software development, inspired by agile methods widely used in professional software engineering.
Its key innovation is a task-oriented approach, where agents aren't given fixed roles but are dynamically assigned tasks during multiple sprint phases, with the task list updated regularly.