Key Points: Stability AI has launched StableLM, its latest open-source model, positioning itself as a contender against giants like OpenAI’s ChatGPT and other alternatives. StableLM showcases the ability to handle various tasks such as code generation and text creation, proving that even compact models can deliver impressive performance with proper training. StableLM is currently in an alpha release stage, accessible on GitHub and Hugging Face platforms.
StableLM:
An Open-Source Challenger to ChatGPT StableLM, the brainchild of Stability AI, emerges as an open-source solution primed for tackling diverse challenges such as content generation and query resolution, placing Stability AI firmly in the competitive ring against OpenAI. As per Stability’s blog, the latest iteration, StableLM, underwent training on an experimental dataset sourced from The Pile, resulting in a dataset three times larger, with approximately 1.5 trillion tokens. Despite its relatively modest parameter count ranging from 3 to 7 billion, StableLM delivers robust performance across coding and conversational tasks, thanks to the richness of its training data. Stability emphasizes in their blog
“Language models form the bedrock of our digital landscape
Through our language model, we strive to empower individuals to express themselves uniquely.” The open-source nature of models like StableLM underscores a commitment to transparency, support, and accessibility in AI technology. Much like OpenAI’s recent GPT-4 release, StableLM excels in generating text and predicting subsequent tokens in a sequence. Users initiate this sequence by providing a prompt or query, and StableLM responds by predicting the next token, capable of producing human-like text and writing code snippets for users. How to Try StableLM Right Now? StableLM is currently available in alpha on GitHub and Hugging Face under the moniker “StableLM-Tuned-Alpha-7b Chat.”
While the Hugging Face version functions akin to ChatGPT,
it may exhibit slower response times compared to other chatbots. Parameter models range from 3 billion to 7 billion, with larger models containing approximately 15 billion and 65 billion parameters in the pipeline. Stability AI affirms, “Our StableLM models possess the capability to generate code and text, enabling a myriad of downstream applications.” This underscores the potential of small yet potent models to deliver exceptional performance with meticulous training. Conclusion In a casual evaluation using the 7B model of StableLM tailored for discussion via the Alpaca method, it surpassed Meta’s raw 7B parameter LLaMA model in output quality, although not reaching the pinnacle of OpenAI’s GPT-3. Nevertheless, the prospect of larger-parameter versions of StableLM suggests enhanced flexibility and efficacy in diverse tasks.