Stability AI, the pioneering generative AI startup behind Stable Diffusion, has unveiled its first Japanese Language Model (LM), known as Japanese StableLM Alpha, in a key step towards improving the Japanese generative AI market.
As the company claims their language model to be the most effective publically available model catering to Japanese speakers, this historic debut has drawn attention. Accordingly to the company, thorough benchmark assessment against four other Japanese LMs supports the assertion. With its design of 7 billion parameters, the recently unveiled Japanese StableLM Alpha is a tribute to Stability AI’s dedication to technological development.
The well-known Apache Licence 2.0 will be used for the commercial distribution of the Japanese StableLM Base Alpha 7B iteration. This specialised model was painstakingly created after prolonged training on a massive dataset that included 750 billion tokens of both Japanese and English text that were carefully collected from web archives.
Read More: OpenAI’s Sam Altman Launches Cryptocurrency Project Worldcoin
The Japanese community of Stability AI created datasets by utilising the knowledge of the EleutherAI Polyglot project’s Japanese team. The use of EleutherAI’s GPT-NeoX software, a key component of Stability AI’s development process, in an expanded form, greatly facilitated this group effort.
The Japanese StableLM Instruct Alpha 7B is a similar model that represents yet another outstanding achievement. This model was created primarily for research purposes and is only suitable for research-related applications. Through the use of several available datasets and a advanced approach known as Supervised Fine-tuning (SFT), it demonstrates a unique capacity to follow user instructions.
EleutherAI’s Language Model Evaluation Harness was used to conduct thorough evaluations that served to validate these models. The models underwent scrutiny across various domains, such as question answering, sentence classification, sentence pair classification, and sentence summarization, emerging with an impressive average score of 54.71%.
According to Stability AI, this performance indicator clearly places the Japanese StableLM Instruct Alpha 7B ahead of its rivals, demonstrating its strength and supremacy.