MediaTek launched the world's first traditional Chinese AI big language model BLOOM-zh, which is now available for use.

CTOnews.com April 30 news, MediaTek April 28 official Xuan, its AI research team recently released the world's first traditional Chinese AI large language model (LLM).

The multilingual model, called BLOOM-zh, claims to outperform its predecessor in most traditional Chinese benchmarks while maintaining its English proficiency.

MediaTek said that BLOOM-zh originated from the open source BLOOM model launched by BigScience in 2022, and the team conducted additional pre-training of 7.4 billion token extensions for BLOOM in traditional Chinese and English, covering news, books, encyclopedias, teaching materials, spoken English and other fields. To demonstrate the features of BLOOM-zh, both existing and newly created benchmark scenarios are used to evaluate performance.

In addition, the AI research group of MediaTek presented a paper at the ICLR International representation Learning Conference, detailing its new optimization algorithm using Fisher-Legendre (FishLeg), which can train the AI model faster and more reliably than before. The study was conducted in collaboration with the Neurodynamics and Control Group at the University of Cambridge in the UK. The MediaTek AI research group will present the results at the ICLR meeting on May 1.

Links to related content are attached to CTOnews.com:

BLOOM-zh thesis: click here to see

BLOOM-zh large model: click here to use

FishLeg algorithm thesis: click here to see

FishLeg open source page: click here to enter