The language technology company Miðeind is featured in a new blog post from OpenAI . A specially selected dataset for the Icelandic language has been incorporated into GPT-4 Turbo, the latest model in the GPT family. The improved model shows significant progress in generating Icelandic text, both in terms of grammar and general fluency.

As part of the work on the project, Miðeind created a benchmark test that evaluates the ability of large language models to correctly inflect Icelandic noun phrases (concordant adjectives and nouns), which previous generations of models have struggled with. The results are very encouraging: For more common words, GPT-4 Turbo inflects entire declension examples (i.e., all cases in singular and plural) flawlessly in 66% of cases, compared to 25% accuracy for the older model released in March.

According to Linda Heimisdóttir, CEO of Miðeind, the company has had a successful collaboration with OpenAI that began following a visit by the President of Iceland, ministers, and a delegation to the company's headquarters in San Francisco last year, where Vilhjálmur Þorsteinsson, founder and chairman of Miðeind, was part of the group.

"OpenAI is a pioneer in the field of large language models and artificial intelligence. This collaboration between Miðeind and OpenAI represents a unique opportunity to advance Icelandic language technology. It's great to see how interested this large company is in the cause of small languages, and we have jointly worked hard to achieve these results. The collaboration continues, and we aim even higher regarding the results."

Hurray for the Icelandic language and happy Icelandic Language Week!