This question is part of the Hill Climbing Round of the Forecasting AI Progress Tournament. You can view all other questions in this round here.

Language modelling has been applied to a wide range of applications and domains with great success. To name a few, automatic speech recognition, machine translation, spelling correction, touchscreen “soft” keyboards and many natural language processing applications depend on the quality of language models.

The One Billion Word dataset, is a large dataset that consists of 829,250,940 tokens over a vocabulary of 793,471 words. Importantly, sentences in this model are shuffled and hence context is limited.

As of writing this question, the state-of-the-art model for is Transformer-XL Dai et al., 2019, which achieves at perplexity of 21.8.

An excellent reference for tracking state-of-the-art models is PapersWithCode, which tracks performance data of ML models.

What will the state-of-the-art language modelling performance on One Billion Word be on 2022-01-14, in perplexity amongst models not trained on additional data?

This question resolves as the lowest level of perplexity achieved by any language model on One Billion Words's test set up until 2022-01-14, 11:59 GMT. Qualifying models need to be trained on only the One Billion Words's training set—no extra training data may be used.