{"about":{"site":"https://codewithpapers.app","non_affiliation":"Code with Papers and Syntology are not affiliated with, endorsed by, or sponsored by Papers with Code, Meta, or the pwc-archive mirror.","licence":"CC BY-SA 4.0","licence_url":"https://creativecommons.org/licenses/by-sa/4.0/legalcode","attribution":"https://codewithpapers.app/attribution","modified":"archive material modified by Syntology; see the attribution page"},"url":"/paper/dynamic-evaluation-of-transformer-language","title":"Dynamic Evaluation of Transformer Language Models","arxiv_id":"1904.08378","date":"2019-04-17","proceeding":null,"authors":["Ben Krause","Emmanuel Kahembwe","Iain Murray","Steve Renals"],"abstract":"This research note combines two methods that have recently improved the state\nof the art in language modeling: Transformers and dynamic evaluation.\nTransformers use stacked layers of self-attention that allow them to capture\nlong range dependencies in sequential data. Dynamic evaluation fits models to\nthe recent sequence history, allowing them to assign higher probabilities to\nre-occurring sequential patterns. By applying dynamic evaluation to\nTransformer-XL models, we improve the state of the art on enwik8 from 0.99 to\n0.94 bits/char, text8 from 1.08 to 1.04 bits/char, and WikiText-103 from 18.3\nto 16.4 perplexity points.","url_abs":"http://arxiv.org/abs/1904.08378v1","url_pdf":"http://arxiv.org/pdf/1904.08378v1.pdf","source":{"archive":"pwc-archive (Hugging Face), CC BY-SA 4.0","snapshot":"2025-07-28","licence_url":"https://creativecommons.org/licenses/by-sa/4.0/legalcode","row_kind":"abstracts"},"code_links":[{"paper_slug":"dynamic-evaluation-of-transformer-language","repo_url":"https://github.com/benkrause/dynamiceval-transformer","is_official":0,"mentioned_in_paper":1,"mentioned_in_github":0,"framework":"tf","reach":null}],"tasks":[{"task_slug":"language-modeling","task_name":"Language Modeling"},{"task_slug":"language-modelling","task_name":"Language Modelling"}],"methods":[],"datasets_introduced":[],"methods_introduced":[],"results":[{"leaderboard":"/sota/language-modelling-on-hutter-prize","task":"Language Modelling","dataset":"Hutter Prize","model":"Transformer-XL + RMS dynamic eval","rank_in_archive_order":1,"of":18,"metrics":{"Bit per Character (BPC)":"0.94","Number of params":"277M"},"uses_additional_data":false},{"leaderboard":"/sota/language-modelling-on-text8","task":"Language Modelling","dataset":"Text8","model":"Transformer-XL + RMS dynamic eval + decay","rank_in_archive_order":3,"of":24,"metrics":{"Bit per Character (BPC)":"1.038","Number of params":"277M"},"uses_additional_data":false},{"leaderboard":"/sota/language-modelling-on-wikitext-103","task":"Language Modelling","dataset":"WikiText-103","model":"Transformer-XL (RMS dynamic eval)","rank_in_archive_order":13,"of":89,"metrics":{"Number of params":"257M","Test perplexity":"16.4","Validation perplexity":"15.8"},"uses_additional_data":true},{"leaderboard":"/sota/language-modelling-on-wikitext-103","task":"Language Modelling","dataset":"WikiText-103","model":"Transformer-XL (SGD dynamic eval)","rank_in_archive_order":19,"of":89,"metrics":{"Number of params":"257M","Test perplexity":"17.0","Validation perplexity":"16.3"},"uses_additional_data":false},{"leaderboard":"/sota/language-modelling-on-enwiki8","task":"Language Modelling","dataset":"enwik8","model":"Transformer-XL (24 layers, RMS dynamic eval, decay)","rank_in_archive_order":2,"of":42,"metrics":{"Bit per Character (BPC)":"0.940","Number of params":"277M"},"uses_additional_data":true}],"syntology":{"atlas_url":"https://app.syntology.ai/?focus=1904.08378","mcp":null,"developers":"https://syntology.ai/developers"},"arxiv_metadata":null,"syntology_extracted_results":null}