Metas Llama and the current Llama 2 are the foundation for two new language models, FreeWilly1 and FreeWilly 2. Published by Stability AI, both models are Open Access and explicitly intended for scientific use. FreeWilly1 is based on the LLaMA 65B Foundation model and has been fine-tuned using a synthetically generated dataset with Supervised Fine-Tune (SFT) in Alpaca format.
On the other hand, FreeWilly 2, which is the more powerful model, is based on Metas Llama 2. It has only recently become available and can be used for commercial purposes, unlike the building LLMs of Stability AI that are specifically excluded from commercial use. The community is encouraged to provide feedback and engage in red teaming to improve security measures.
To fine-tune the language models, Stability AI utilized the Orca method developed by Microsoft. This method involves a smaller model learning the reasoning style of a larger one instead of simply imitating the output. The training set used for this process is available from GPT-4. While the trained smaller model may not be better or as good as the larger model, it has the potential to outperform other models of a similar size. Stability AI has also released benchmarks, showcasing the performance of FreeWilly and FreeWilly2 in logical tasks. The blog post emphasizes that both models set a new standard for open access of Large Language Models, significantly contributing to research, improving natural language understanding, and enabling complex tasks.
In addition to the well-known Stable Diffusion, Stability AI offers other image generators in its portfolio, namely Stable Doodle and Stable Studio. Each of these generators specializes in sketches or graphics. For creating particularly realistic images, there is Stable Diffusion XL 0.9.
The article concludes by providing a link to the home page of Stability AI.
