Meta stands up to OpenAI and Google with LIMA, its new AI model

May 24. (Portaltic/EP) –

Goal has presented LIMEa big language model with which they demonstrate that high-quality responses can be obtained from a small set of indications with a previously trained model.

LIMA (Less Is More for Alignment)sIt is based on LLaMa, a model with 65,000 million parameters that the technology company provided for research purposes in the first quarter of the year.

Meta explains that large language models are usually trained in two phases: an unsupervised pre-training of raw text, so that it learns general representations, and a large-scale one of adjustment and reinforcement learning, with which it is intended that the AI better aligns with end tasks and user preferences.

With LIMA, Meta aims to demonstrate that it is possible to obtain quality results from a few indications with a model that has been extensively trained before. And for this, he has used thousand examples of real instructions carefully curated, 750 from forums like Stack Exchange and wikiHow and another 250 written by the researchers themselves.

To analyze its performance, they have compared it with OpenAI GPT-4Claude by Anthropic and bard from google with a controlled test of 300 indications. The results they obtained show that LIMA produces “same or preferable” responses in 43 percent, 46 percent, and 58 percent of cases, respectively.

As they collect in the published study at arxiv.orgon an absolute scale, LIMA’s responses “reveal that 88 percent meet the immediate requirements, and 50 percent are considered excellent”the researchers point out.