Although Stability AI is primarily recognized for its robust diffusion text-to-image generative AI models, the company’s scope of operations has expanded significantly.
The latest iteration of Stability AI’s StableLM Zephyr 3B, a 3 billion parameter large language model (LLM) tailored for chat applications such as text generation, summarization, and content personalization, was officially launched today. Initially introduced in April, the StableLM word generation model has undergone enhancements, resulting in a more compact and efficient version.
StableLM Zephyr 3B boasts several advantages, including its smaller size compared to the 7 billion STABLELM variants. This reduction in size enables easier deployment, resource optimization, and swift responsiveness, particularly optimized for applications like Q&A and educational purposes.
Emad Mostaque, the CEO of Stability AI, highlighted that the training process for StableLM involved superior quality data and extended duration compared to previous models. Despite being 40% smaller in scale, it matches the basic functionality of the LLaMA v2 7b model, with twice the token count.
What is the objective of StableLM Zephyr 3B?
Stability AI characterizes StableLM Zephyr 3B as an evolution of the existing 4e1t design rather than an entirely new creation.
The design methodology for Zephyr drew inspiration from the Zephyr 7B model by HuggingFace, renowned for its assistant capabilities and open-source MIT registration. Direct Preference Optimization (DPO), a training approach utilized by Zephyr, has now been adopted by StableLM, offering significant benefits.
Mostaque explained that DPO represents a departure from earlier reinforcement learning techniques, focusing on aligning models with human preferences. StableLM Zephyr, with its 3 billion parameter size, was among the first to implement DPO, typically associated with larger 7 billion parameter models.
Stability AI leveraged DPO in conjunction with the UltraFeedback dataset from the OpenBMB research team, comprising over 256,000 responses and 64,000 prompts. The tailored instruction set, smaller dimension, and DPO integration contributed to StableLM Zephyr 3B’s impressive performance metrics, surpassing more complex models in evaluations like Meta’s Llama 270b talk and Anthropric Claude V1 on the MT Bench analysis.
Diversification of Stability AI’s Product Portfolio
Amid its continuous innovation efforts, Stability AI has introduced a series of new products, with StableLM Zephyr 3B being the latest addition to its expanding lineup.
In August, Stability AI unveiled StableCode, an IoT model designed for software code generation. Subsequently, in September, the company launched Secure Audio, a novel text-to-audio synthesis tool. November saw Stability AI’s foray into the video technology domain with the introduction of Stable Video Diffusion.
Despite these new ventures, Stability AI remains committed to its text-to-image technology roots, with upcoming releases poised to further enhance its offerings in this space.
Mostaque emphasized the potential of smaller, efficient models tailored to individual user data, suggesting their potential to outperform larger public models. Anticipating the forthcoming release of new StableLM models, Stability AI aims to empower users with advanced relational language models.
