Apple has released OpenELM (Open-source Efficient Language Models), a family of eight open-source large language models (LLMs) on Hugging Face. The models cover a parameter range of 270 million to 3 billion parameters and outperform Allen Institute for AI’s OLMo Models by 2.36% while requiring fewer pre-training tokens. OpenELM includes four pre-trained models and four instruct models, with data training from publicly available sources like Wikipedia, Reddit, and Project Gutenberg. Apple’s release of OpenELM includes the complete framework for training and evaluation on publicly available datasets, making it different from Meta’s open-source releases that have restrictions baked into the license. Apple has not released any commercial models but has been expanding its AI businesses through acquisitions and collaborations, such as acquiring Darwin AI and potentially incorporating Google’s Gemini into iPhones. Users are allowed to use, reproduce, modify, and redistribute the OpenELM models under the Apple Sample Code License. This move by Apple signifies a step towards openness in the AI space, contrasting with Meta’s restrictions on commercial usage in their open-source releases.
Source link
Source link: https://www.medianama.com/2024/04/223-apple-family-open-source-large-language-models-openelm/
Introducing a New Family of AI Models #Innovation
Preeti
1.5k Views
0 Votes
GIPHY App Key not set. Please check settings