{"url":"/sota/language-modelling-on-one-billion-word","task":{"name":"Language Modelling","url":"/task/language-modelling","note":null},"dataset":{"name":"One Billion Word","url":"/dataset/billion-word-benchmark"},"category":"Natural Language Processing","categories":["Medical","Miscellaneous","Natural Language Processing"],"category_note":null,"description":"A language model is a model of natural language. Language models are useful for a variety of tasks, including speech recognition, machine translation, natural language generation (generating more human-like text), optical character recognition, route optimization, handwriting recognition, grammar induction, and information retrieval.\r\n\r\nLarge language models (LLMs), currently their most advanced form, are predominantly based on transformers trained on larger datasets (frequently using words scraped from the public internet). They have superseded recurrent neural network-based models, which had previously superseded the purely statistical models, such as word n-gram language model. \r\n\r\nSource: [Wikipedia](https://en.wikipedia.org/wiki/Language_model)","description_from":"task","source":{"archive":"pwc-archive (Hugging Face), CC BY-SA 4.0","snapshot":"2025-07-28","rank":"the archive's row order at snapshot; not re-ranked","rows_end_at":"2025-07-28","rows_withheld_as_spam":0,"metric_values":"the archive's strings, untouched"},"metrics":["PPL","Number of params","Validation perplexity"],"metric_direction":{"note":"inferred from the metric name only (the archive records no direction); null = not inferred, chart draws points only","by_metric":{"PPL":"lower","Number of params":"lower","Validation perplexity":"lower"}},"counts":{"rows":27,"rows_with_code":25,"rows_with_paper_page":26,"rows_dated":26,"rows_using_additional_data":1},"rows":[{"rank_in_archive_order":1,"model":"MDLM (AR baseline)","metrics":{"Number of params":"110M","PPL":"20.09"},"uses_additional_data":false,"paper_date":"2024-06-11","paper":"/paper/simple-and-effective-masked-diffusion","paper_url":"https://arxiv.org/abs/2406.07524v2","paper_title":"Simple and Effective Masked Diffusion Language Models","code":"https://github.com/kuleshov-group/mdlm","n_code_links":2,"syntology":{"n_ran":1,"n_unverified":0,"n_samples":1,"n_pointer_only_licence":1}},{"rank_in_archive_order":2,"model":"OmniNetT (Large)","metrics":{"Number of params":"100M","PPL":"21.5"},"uses_additional_data":false,"paper_date":"2021-03-01","paper":"/paper/omninet-omnidirectional-representations-from","paper_url":"https://arxiv.org/abs/2103.01075v1","paper_title":"OmniNet: Omnidirectional Representations from Transformers","code":"https://github.com/lucidrains/omninet-pytorch","n_code_links":1,"syntology":null},{"rank_in_archive_order":3,"model":"OmniNetP (Large)","metrics":{"Number of params":"100M","PPL":"21.6"},"uses_additional_data":false,"paper_date":"2021-03-01","paper":"/paper/omninet-omnidirectional-representations-from","paper_url":"https://arxiv.org/abs/2103.01075v1","paper_title":"OmniNet: Omnidirectional Representations from Transformers","code":"https://github.com/lucidrains/omninet-pytorch","n_code_links":1,"syntology":null},{"rank_in_archive_order":4,"model":"Transformer-XL Large","metrics":{"Number of params":"0.8B","PPL":"21.8"},"uses_additional_data":false,"paper_date":"2019-01-09","paper":"/paper/transformer-xl-attentive-language-models","paper_url":"https://arxiv.org/abs/1901.02860v3","paper_title":"Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context","code":"https://github.com/huggingface/transformers","n_code_links":37,"syntology":{"n_ran":63,"n_unverified":80,"n_samples":143,"n_pointer_only_licence":43}},{"rank_in_archive_order":5,"model":"OmniNetB (Large)","metrics":{"PPL":"22"},"uses_additional_data":false,"paper_date":"2021-03-01","paper":"/paper/omninet-omnidirectional-representations-from","paper_url":"https://arxiv.org/abs/2103.01075v1","paper_title":"OmniNet: Omnidirectional Representations from Transformers","code":"https://github.com/lucidrains/omninet-pytorch","n_code_links":1,"syntology":null},{"rank_in_archive_order":6,"model":"MDLM","metrics":{"Number of params":"110M","PPL":"23.00"},"uses_additional_data":false,"paper_date":"2024-06-11","paper":"/paper/simple-and-effective-masked-diffusion","paper_url":"https://arxiv.org/abs/2406.07524v2","paper_title":"Simple and Effective Masked Diffusion Language Models","code":"https://github.com/kuleshov-group/mdlm","n_code_links":2,"syntology":{"n_ran":1,"n_unverified":0,"n_samples":1,"n_pointer_only_licence":1}},{"rank_in_archive_order":7,"model":"Adaptive Input Very Large","metrics":{"Number of params":"1.0B","PPL":"23.02","Validation perplexity":"22.92"},"uses_additional_data":false,"paper_date":"2018-09-28","paper":"/paper/adaptive-input-representations-for-neural","paper_url":"http://arxiv.org/abs/1809.10853v3","paper_title":"Adaptive Input Representations for Neural Language Modeling","code":"https://github.com/pytorch/fairseq","n_code_links":3,"syntology":null},{"rank_in_archive_order":8,"model":"Transformer-XL Base","metrics":{"Number of params":"0.46B","PPL":"23.5"},"uses_additional_data":false,"paper_date":"2019-01-09","paper":"/paper/transformer-xl-attentive-language-models","paper_url":"https://arxiv.org/abs/1901.02860v3","paper_title":"Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context","code":"https://github.com/huggingface/transformers","n_code_links":37,"syntology":{"n_ran":63,"n_unverified":80,"n_samples":143,"n_pointer_only_licence":43}},{"rank_in_archive_order":9,"model":"SRU++ Large","metrics":{"Number of params":"465M","PPL":"23.5"},"uses_additional_data":false,"paper_date":"2021-02-24","paper":"/paper/when-attention-meets-fast-recurrence-training","paper_url":"https://arxiv.org/abs/2102.12459v3","paper_title":"When Attention Meets Fast Recurrence: Training Language Models with Reduced Compute","code":"https://github.com/asappresearch/sru","n_code_links":1,"syntology":null},{"rank_in_archive_order":10,"model":"10 LSTM+CNN inputs + SNM10-SKIP (ensemble)","metrics":{"Number of params":"43B","PPL":"23.7"},"uses_additional_data":false,"paper_date":"2016-02-07","paper":"/paper/exploring-the-limits-of-language-modeling","paper_url":"http://arxiv.org/abs/1602.02410v2","paper_title":"Exploring the Limits of Language Modeling","code":"https://github.com/tensorflow/models","n_code_links":10,"syntology":{"n_ran":2,"n_unverified":6,"n_samples":8,"n_pointer_only_licence":0}},{"rank_in_archive_order":11,"model":"Adaptive Input Large","metrics":{"Number of params":"0.46B","PPL":"23.91","Validation perplexity":"23.83"},"uses_additional_data":false,"paper_date":"2018-09-28","paper":"/paper/adaptive-input-representations-for-neural","paper_url":"http://arxiv.org/abs/1809.10853v3","paper_title":"Adaptive Input Representations for Neural Language Modeling","code":"https://github.com/pytorch/fairseq","n_code_links":3,"syntology":null},{"rank_in_archive_order":12,"model":"Mesh Tensorflow","metrics":{"Number of params":"4.9B","PPL":" 24.0"},"uses_additional_data":false,"paper_date":"2018-11-05","paper":"/paper/mesh-tensorflow-deep-learning-for","paper_url":"http://arxiv.org/abs/1811.02084v1","paper_title":"Mesh-TensorFlow: Deep Learning for Supercomputers","code":"https://github.com/tensorflow/mesh","n_code_links":1,"syntology":null},{"rank_in_archive_order":13,"model":"Cohere Large","metrics":{"PPL":"25.06"},"uses_additional_data":false,"paper_date":null,"paper":null,"paper_url":null,"paper_title":"","code":null,"n_code_links":0,"syntology":null},{"rank_in_archive_order":14,"model":"SRU++","metrics":{"Number of params":"328M","PPL":"25.1"},"uses_additional_data":false,"paper_date":"2021-02-24","paper":"/paper/when-attention-meets-fast-recurrence-training","paper_url":"https://arxiv.org/abs/2102.12459v3","paper_title":"When Attention Meets Fast Recurrence: Training Language Models with Reduced Compute","code":"https://github.com/asappresearch/sru","n_code_links":1,"syntology":null},{"rank_in_archive_order":15,"model":"DynamicConv","metrics":{"Number of params":"0.34B","PPL":"26.67"},"uses_additional_data":false,"paper_date":"2019-01-29","paper":"/paper/pay-less-attention-with-lightweight-and","paper_url":"http://arxiv.org/abs/1901.10430v2","paper_title":"Pay Less Attention with Lightweight and Dynamic Convolutions","code":"https://github.com/pytorch/fairseq","n_code_links":3,"syntology":null},{"rank_in_archive_order":16,"model":"High-Budget MoE","metrics":{"Number of params":"5B","PPL":" 28.0"},"uses_additional_data":false,"paper_date":"2017-01-23","paper":"/paper/outrageously-large-neural-networks-the","paper_url":"http://arxiv.org/abs/1701.06538v1","paper_title":"Outrageously Large Neural Networks: The Sparsely-Gated Mixture-of-Experts Layer","code":"https://github.com/davidmrau/mixture-of-experts","n_code_links":4,"syntology":{"n_ran":4,"n_unverified":2,"n_samples":6,"n_pointer_only_licence":6}},{"rank_in_archive_order":17,"model":"Evolved Transformer Big","metrics":{"PPL":"28.6"},"uses_additional_data":false,"paper_date":"2019-01-30","paper":"/paper/the-evolved-transformer","paper_url":"https://arxiv.org/abs/1901.11117v4","paper_title":"The Evolved Transformer","code":"https://github.com/tensorflow/tensor2tensor","n_code_links":3,"syntology":null},{"rank_in_archive_order":18,"model":"LSTM-8192-1024 + CNN Input","metrics":{"Number of params":"1.04B","PPL":"30.0"},"uses_additional_data":false,"paper_date":"2016-02-07","paper":"/paper/exploring-the-limits-of-language-modeling","paper_url":"http://arxiv.org/abs/1602.02410v2","paper_title":"Exploring the Limits of Language Modeling","code":"https://github.com/tensorflow/models","n_code_links":10,"syntology":{"n_ran":2,"n_unverified":6,"n_samples":8,"n_pointer_only_licence":0}},{"rank_in_archive_order":19,"model":"LSTM-8192-1024","metrics":{"Number of params":"1.8B","PPL":"30.6"},"uses_additional_data":false,"paper_date":"2016-02-07","paper":"/paper/exploring-the-limits-of-language-modeling","paper_url":"http://arxiv.org/abs/1602.02410v2","paper_title":"Exploring the Limits of Language Modeling","code":"https://github.com/tensorflow/models","n_code_links":10,"syntology":{"n_ran":2,"n_unverified":6,"n_samples":8,"n_pointer_only_licence":0}},{"rank_in_archive_order":20,"model":"GCNN-14 bottleneck","metrics":{"PPL":"31.9"},"uses_additional_data":false,"paper_date":"2016-12-23","paper":"/paper/language-modeling-with-gated-convolutional","paper_url":"http://arxiv.org/abs/1612.08083v3","paper_title":"Language Modeling with Gated Convolutional Networks","code":"https://github.com/facebookresearch/fairseq","n_code_links":11,"syntology":null},{"rank_in_archive_order":21,"model":"Low-Budget MoE","metrics":{"Number of params":"5B","PPL":"34.1"},"uses_additional_data":false,"paper_date":"2017-01-23","paper":"/paper/outrageously-large-neural-networks-the","paper_url":"http://arxiv.org/abs/1701.06538v1","paper_title":"Outrageously Large Neural Networks: The Sparsely-Gated Mixture-of-Experts Layer","code":"https://github.com/davidmrau/mixture-of-experts","n_code_links":4,"syntology":{"n_ran":4,"n_unverified":2,"n_samples":6,"n_pointer_only_licence":6}},{"rank_in_archive_order":22,"model":"BIG G-LSTM-2","metrics":{"PPL":"36.0"},"uses_additional_data":false,"paper_date":"2017-03-31","paper":"/paper/factorization-tricks-for-lstm-networks","paper_url":"http://arxiv.org/abs/1703.10722v3","paper_title":"Factorization tricks for LSTM networks","code":"https://github.com/okuchaiev/f-lm","n_code_links":2,"syntology":null},{"rank_in_archive_order":23,"model":"GPT-2","metrics":{"Number of params":"1.54B","PPL":"42.16"},"uses_additional_data":true,"paper_date":"2019-02-14","paper":"/paper/language-models-are-unsupervised-multitask","paper_url":"https://d4mucfpksywv.cloudfront.net/better-language-models/language-models.pdf","paper_title":"Language Models are Unsupervised Multitask Learners","code":"https://github.com/huggingface/transformers","n_code_links":21,"syntology":null},{"rank_in_archive_order":24,"model":"RNN-1024 + 9 Gram","metrics":{"Number of params":"20B","PPL":"51.3"},"uses_additional_data":false,"paper_date":"2013-12-11","paper":"/paper/one-billion-word-benchmark-for-measuring","paper_url":"http://arxiv.org/abs/1312.3005v3","paper_title":"One Billion Word Benchmark for Measuring Progress in Statistical Language Modeling","code":"https://github.com/tensorflow/models","n_code_links":3,"syntology":null},{"rank_in_archive_order":25,"model":"Sparse Non-Negative","metrics":{"Number of params":"33B","PPL":"52.9"},"uses_additional_data":false,"paper_date":"2014-12-03","paper":"/paper/skip-gram-language-modeling-using-sparse-non","paper_url":"http://arxiv.org/abs/1412.1454v2","paper_title":"Skip-gram Language Modeling Using Sparse Non-negative Matrix Probability Estimation","code":null,"n_code_links":0,"syntology":null},{"rank_in_archive_order":26,"model":"H-Transformer-1D Nr=16 (Base)","metrics":{"Number of params":"53M","Validation perplexity":"23.95"},"uses_additional_data":false,"paper_date":"2021-07-25","paper":"/paper/h-transformer-1d-fast-one-dimensional","paper_url":"https://arxiv.org/abs/2107.11906v1","paper_title":"H-Transformer-1D: Fast One-Dimensional Hierarchical Attention for Sequences","code":"https://github.com/lucidrains/h-transformer-1d","n_code_links":2,"syntology":{"n_ran":5,"n_unverified":5,"n_samples":10,"n_pointer_only_licence":0}},{"rank_in_archive_order":27,"model":"H-Transformer-1D Nr=16 (Large)","metrics":{"Number of params":"144M","Validation perplexity":"20.25"},"uses_additional_data":false,"paper_date":"2021-07-25","paper":"/paper/h-transformer-1d-fast-one-dimensional","paper_url":"https://arxiv.org/abs/2107.11906v1","paper_title":"H-Transformer-1D: Fast One-Dimensional Hierarchical Attention for Sequences","code":"https://github.com/lucidrains/h-transformer-1d","n_code_links":2,"syntology":{"n_ran":5,"n_unverified":5,"n_samples":10,"n_pointer_only_licence":0}}],"since_archive":{"claim":"Results that newer papers report for their own method, placed here by Syntology. A model pointed at the cell in the paper's own table; the number was read from that cell and checked against this leaderboard's metric, dataset, split and scale; an independent check that saw this leaderboard's other rows and every other leaderboard on the same dataset accepted it. Not reviewed by the paper's authors or by the archive's editors, and not ranked against the archive rows.","extraction_file_present":true,"measurement":{"test_papers":883,"papers_with_output":881,"judged_true":108,"judged":110,"wilson95_lower":0.9361,"measured_on":"2026-09-24","frozen_commit":"0e3de0df94"},"measurement_note":"blind adjudication of accepted entries on a held-out split of archive papers, rules frozen before the test","coverage":{"sentence":"Syntology has checked 6,264 of the 9,581 papers on this site that are newer than the archive; results from the others appear after they are checked.","complete":false,"papers_newer_than_archive":9581,"papers_checked":6264,"papers_extracted_not_yet_verified":0,"boards_without_verdict":2,"papers_not_yet_extracted":3316},"order":"newest first by month (arXiv date, else the arXiv-id month), then arXiv id descending","columns":[],"entries":[]},"syntology":{"read_at":"2026-09-24T18:15:14+00:00","claim":"Per row: N of M harvested code samples from that row's paper executed on a synthesized fixture; the other M-N are unverified. Not a reproduction of the row's number; not a correctness claim. n_pointer_only_licence counts samples the site points at rather than redistributes (a licence axis, independent of ran/unverified).","rows_with_graph_line":11,"rows_with_any_sample_ran":11,"distinct_papers_with_graph_line":5,"distinct_papers_with_any_sample_ran":5,"samples_over_distinct_papers":{"n_ran":75,"n_unverified":93,"n_samples":168,"n_pointer_only_licence":50,"note":"each paper (arXiv id) counted once, however many rows it is behind; this is the page-level figure"},"samples_row_weighted":{"n_ran":152,"n_unverified":192,"n_samples":344,"n_pointer_only_licence":100,"note":"row-weighted: a paper behind several rows is counted once per row; inflated relative to samples_over_distinct_papers by design, kept for readers summing the per-row syntology blocks"}}}