MADRID–(BUSINESS WIRE)–August 4, 2023–
Clibrain announces a milestone in artificial intelligence: the creation of its first language model (LLM) fully adapted and trained in Spanish: LINCE. This model is an unprecedented achievement in our language and it’s just the start of a long journey. In this way, Clibrain is at the forefront of the evolution of AI, taking a step forward on the path to a future where communication between humans and machines is reaching its peak expression.
The development of this model is an innovation in the training process, corpus, and techniques employed to achieve the best performance, applying the latest efficient fine-tuning techniques on LLMs, the result of a completely new and previously non-existent dataset. This model is made with an instruction-focused approach.
Why does a language model trained entirely in Spanish emerge?
In the globalized world in which we live, there is a range of more than 21 variations of Spanish for which there are no specialized language models. Machines typically offer two or three varieties, among which are usually Spanish from Spain and “Latin,” but without much more variety. And more importantly, in general, the responses and instructions in our language – like in the rest – are the result of translating from English to other languages, which implies a loss of quality and nuances that puts us at a disadvantage from the outset.
“If the context for generative AI is important, the cultural context is even more so, and especially when we talk about language varieties and dialects”, explains Elena González-Blanco, CEO of Clibrain.
This challenge becomes especially noticeable in disciplines with technical and specialized language such as legal, financial, or medical, as well as in the interpretation of proverbs, idiomatic phrases, and even in the analysis of old Spanish.
An Open Source model
Another interesting aspect of the journey Clibrain undertakes is the launch of two versions of the language model: the first one is already released under an open source license, featuring more than 7 billion parameters called “LINCE ZERO”, which can be used by anyone to create their applications or utilize it for non-commercial purposes; the final version, “LINCE“, will be launched in the coming weeks.
Join the waiting list
If you want to be one of the first to test LINCE, you can request access at email@example.com.
VentureBeat's mission is to be a digital town square for technical decision-makers to gain knowledge about transformative enterprise technology and transact. Discover our Briefings.