{"url":"/sota/language-modelling-on-wikitext-2","task":{"name":"Language Modelling","url":"/task/language-modelling","note":null},"dataset":{"name":"WikiText-2","url":"/dataset/wikitext-2"},"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 They have superseded recurrent neural network-based models, which had previously superseded the purely statistical models, such as word n-gram language model.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":["Test perplexity","Validation perplexity","Number of params"],"metric_direction":{"note":"inferred from the metric name only (the archive records no direction); null = not inferred, chart draws points only","by_metric":{"Test perplexity":"lower","Validation perplexity":"lower","Number of params":"lower"}},"counts":{"rows":38,"rows_with_code":37,"rows_with_paper_page":38,"rows_dated":38,"rows_using_additional_data":11},"rows":[{"rank_in_archive_order":1,"model":"SparseGPT (175B, 50% Sparsity)","metrics":{"Test perplexity":"8.21"},"uses_additional_data":true,"paper_date":"2023-01-02","paper":"/paper/massive-language-models-can-be-accurately","paper_url":"https://arxiv.org/abs/2301.00774v3","paper_title":"SparseGPT: Massive Language Models Can Be Accurately Pruned in One-Shot","code":"https://github.com/nvidia/tensorrt-model-optimizer","n_code_links":6,"syntology":{"n_ran":2,"n_unverified":10,"n_samples":12,"n_pointer_only_licence":9}},{"rank_in_archive_order":2,"model":"OPT-175B","metrics":{"Test perplexity":"8.34"},"uses_additional_data":true,"paper_date":"2023-01-02","paper":"/paper/massive-language-models-can-be-accurately","paper_url":"https://arxiv.org/abs/2301.00774v3","paper_title":"SparseGPT: Massive Language Models Can Be Accurately Pruned in One-Shot","code":"https://github.com/nvidia/tensorrt-model-optimizer","n_code_links":6,"syntology":{"n_ran":2,"n_unverified":10,"n_samples":12,"n_pointer_only_licence":9}},{"rank_in_archive_order":3,"model":"SparseGPT (175B, 4:8 Sparsity)","metrics":{"Test perplexity":"8.45"},"uses_additional_data":true,"paper_date":"2023-01-02","paper":"/paper/massive-language-models-can-be-accurately","paper_url":"https://arxiv.org/abs/2301.00774v3","paper_title":"SparseGPT: Massive Language Models Can Be Accurately Pruned in One-Shot","code":"https://github.com/nvidia/tensorrt-model-optimizer","n_code_links":6,"syntology":{"n_ran":2,"n_unverified":10,"n_samples":12,"n_pointer_only_licence":9}},{"rank_in_archive_order":4,"model":"SparseGPT (175B, 2:4 Sparsity)","metrics":{"Test perplexity":"8.73"},"uses_additional_data":true,"paper_date":"2023-01-02","paper":"/paper/massive-language-models-can-be-accurately","paper_url":"https://arxiv.org/abs/2301.00774v3","paper_title":"SparseGPT: Massive Language Models Can Be Accurately Pruned in One-Shot","code":"https://github.com/nvidia/tensorrt-model-optimizer","n_code_links":6,"syntology":{"n_ran":2,"n_unverified":10,"n_samples":12,"n_pointer_only_licence":9}},{"rank_in_archive_order":5,"model":"GPT-2 (fine-tuned)","metrics":{"Number of params":"1542M","Test perplexity":"15.17","Validation perplexity":"15.69"},"uses_additional_data":true,"paper_date":"2021-10-16","paper":"/paper/hydra-a-system-for-large-multi-model-deep","paper_url":"https://arxiv.org/abs/2110.08633v7","paper_title":"Hydra: A System for Large Multi-Model Deep Learning","code":"https://github.com/knagrecha/hydra","n_code_links":1,"syntology":null},{"rank_in_archive_order":6,"model":"GPT-2","metrics":{"Number of params":"1542M","Test perplexity":"18.34"},"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":7,"model":"GPT-2 (large)","metrics":{"Number of params":"762M","Test perplexity":"19.93"},"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":8,"model":"GPT-2 (medium)","metrics":{"Number of params":"345M","Test perplexity":"22.76"},"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":9,"model":"GPT-2 (small)","metrics":{"Number of params":"117M","Test perplexity":"29.41"},"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":10,"model":"BERT-Large-CAS","metrics":{"Number of params":"395M","Test perplexity":"34.1","Validation perplexity":"37.7"},"uses_additional_data":true,"paper_date":"2019-04-20","paper":"/paper/190409408","paper_url":"https://arxiv.org/abs/1904.09408v2","paper_title":"Language Models with Transformers","code":"https://github.com/cgraywang/gluon-nlp-1","n_code_links":1,"syntology":{"n_ran":2,"n_unverified":0,"n_samples":2,"n_pointer_only_licence":2}},{"rank_in_archive_order":11,"model":"Mogrifier LSTM + dynamic eval","metrics":{"Number of params":"35M","Test perplexity":"38.6","Validation perplexity":"40.2"},"uses_additional_data":false,"paper_date":"2019-09-04","paper":"/paper/mogrifier-lstm","paper_url":"https://arxiv.org/abs/1909.01792v2","paper_title":"Mogrifier LSTM","code":"https://github.com/deepmind/lamb","n_code_links":3,"syntology":null},{"rank_in_archive_order":12,"model":"adversarial + AWD-LSTM-MoS + dynamic eval","metrics":{"Number of params":"35M","Test perplexity":"38.65","Validation perplexity":"40.27"},"uses_additional_data":false,"paper_date":"2019-06-10","paper":"/paper/improving-neural-language-modeling-via","paper_url":"https://arxiv.org/abs/1906.03805v2","paper_title":"Improving Neural Language Modeling via Adversarial Training","code":"https://github.com/ChengyueGongR/advsoft","n_code_links":1,"syntology":null},{"rank_in_archive_order":13,"model":"FRAGE + AWD-LSTM-MoS + dynamic eval","metrics":{"Number of params":"35M","Test perplexity":"39.14","Validation perplexity":"40.85"},"uses_additional_data":false,"paper_date":"2018-09-18","paper":"/paper/frage-frequency-agnostic-word-representation","paper_url":"https://arxiv.org/abs/1809.06858v2","paper_title":"FRAGE: Frequency-Agnostic Word Representation","code":"https://github.com/ChengyueGongR/FrequencyAgnostic","n_code_links":2,"syntology":null},{"rank_in_archive_order":14,"model":"Past Decode Reg. + AWD-LSTM-MoS + dyn. eval.","metrics":{"Number of params":"35M","Test perplexity":"40.3","Validation perplexity":"42.0"},"uses_additional_data":false,"paper_date":"2018-08-14","paper":"/paper/improved-language-modeling-by-decoding-the","paper_url":"http://arxiv.org/abs/1808.05908v4","paper_title":"Improved Language Modeling by Decoding the Past","code":null,"n_code_links":0,"syntology":null},{"rank_in_archive_order":15,"model":"GL-LWGC + AWD-MoS-LSTM + dynamic eval","metrics":{"Number of params":"38M","Test perplexity":"40.46","Validation perplexity":"42.19"},"uses_additional_data":false,"paper_date":"2017-08-29","paper":"/paper/gradual-learning-of-recurrent-neural-networks","paper_url":"http://arxiv.org/abs/1708.08863v2","paper_title":"Gradual Learning of Recurrent Neural Networks","code":"https://github.com/zivaharoni/gradual-learning-rnn","n_code_links":1,"syntology":null},{"rank_in_archive_order":16,"model":"AWD-LSTM-MoS + dynamic eval","metrics":{"Number of params":"35M","Test perplexity":"40.68","Validation perplexity":"42.41"},"uses_additional_data":false,"paper_date":"2017-11-10","paper":"/paper/breaking-the-softmax-bottleneck-a-high-rank","paper_url":"http://arxiv.org/abs/1711.03953v4","paper_title":"Breaking the Softmax Bottleneck: A High-Rank RNN Language Model","code":"https://github.com/zihangdai/mos","n_code_links":9,"syntology":{"n_ran":1,"n_unverified":22,"n_samples":23,"n_pointer_only_licence":0}},{"rank_in_archive_order":17,"model":"AWD-LSTM-DRILL + dynamic eval","metrics":{"Number of params":"34M","Test perplexity":"42.0","Validation perplexity":"43.9"},"uses_additional_data":false,"paper_date":"2019-05-14","paper":"/paper/deep-residual-output-layers-for-neural","paper_url":"https://arxiv.org/abs/1905.05513v2","paper_title":"Deep Residual Output Layers for Neural Language Generation","code":"https://github.com/idiap/drill","n_code_links":1,"syntology":null},{"rank_in_archive_order":18,"model":"AWD-LSTM + dynamic eval","metrics":{"Number of params":"33M","Test perplexity":"44.3","Validation perplexity":"46.4"},"uses_additional_data":false,"paper_date":"2017-09-21","paper":"/paper/dynamic-evaluation-of-neural-sequence-models","paper_url":"http://arxiv.org/abs/1709.07432v2","paper_title":"Dynamic Evaluation of Neural Sequence Models","code":"https://github.com/benkrause/dynamic-evaluation","n_code_links":3,"syntology":{"n_ran":0,"n_unverified":1,"n_samples":1,"n_pointer_only_licence":0}},{"rank_in_archive_order":19,"model":"AWD-LSTM + continuous cache pointer","metrics":{"Number of params":"33M","Test perplexity":"52.0","Validation perplexity":"53.8"},"uses_additional_data":false,"paper_date":"2017-08-07","paper":"/paper/regularizing-and-optimizing-lstm-language","paper_url":"http://arxiv.org/abs/1708.02182v1","paper_title":"Regularizing and Optimizing LSTM Language Models","code":"https://github.com/google-research/google-research/tree/master/enas_lm","n_code_links":45,"syntology":{"n_ran":6,"n_unverified":1,"n_samples":7,"n_pointer_only_licence":7}},{"rank_in_archive_order":20,"model":"AWD-LSTM-DOC x5","metrics":{"Number of params":"185M","Test perplexity":"53.09","Validation perplexity":"54.19"},"uses_additional_data":false,"paper_date":"2018-08-30","paper":"/paper/direct-output-connection-for-a-high-rank","paper_url":"http://arxiv.org/abs/1808.10143v2","paper_title":"Direct Output Connection for a High-Rank Language Model","code":"https://github.com/nttcslab-nlp/doc_lm","n_code_links":1,"syntology":null},{"rank_in_archive_order":21,"model":"Ensemble of All","metrics":{"Test perplexity":"53.73","Validation perplexity":"55.4"},"uses_additional_data":false,"paper_date":"2023-11-28","paper":"/paper/advancing-state-of-the-art-in-language","paper_url":"https://arxiv.org/abs/2312.03735v1","paper_title":"Advancing State of the Art in Language Modeling","code":"https://github.com/davidherel/sota_lm","n_code_links":1,"syntology":null},{"rank_in_archive_order":22,"model":"Mogrifier LSTM","metrics":{"Number of params":"35M","Test perplexity":"55.1","Validation perplexity":"57.3"},"uses_additional_data":false,"paper_date":"2019-09-04","paper":"/paper/mogrifier-lstm","paper_url":"https://arxiv.org/abs/1909.01792v2","paper_title":"Mogrifier LSTM","code":"https://github.com/deepmind/lamb","n_code_links":3,"syntology":null},{"rank_in_archive_order":23,"model":"AWD-LSTM-DOC + Partial Shuffle","metrics":{"Number of params":"37M","Test perplexity":"57.85","Validation perplexity":"60.16"},"uses_additional_data":false,"paper_date":"2019-03-11","paper":"/paper/partially-shuffling-the-training-data-to-1","paper_url":"http://arxiv.org/abs/1903.04167v2","paper_title":"Partially Shuffling the Training Data to Improve Language Models","code":"https://github.com/ofirpress/PartialShuffle","n_code_links":1,"syntology":null},{"rank_in_archive_order":24,"model":"AWD-LSTM-DOC","metrics":{"Number of params":"37M","Test perplexity":"58.03","Validation perplexity":"60.29"},"uses_additional_data":false,"paper_date":"2018-08-30","paper":"/paper/direct-output-connection-for-a-high-rank","paper_url":"http://arxiv.org/abs/1808.10143v2","paper_title":"Direct Output Connection for a High-Rank Language Model","code":"https://github.com/nttcslab-nlp/doc_lm","n_code_links":1,"syntology":null},{"rank_in_archive_order":25,"model":"AWD-LSTM-MoS + Partial Shuffle","metrics":{"Number of params":"35M","Test perplexity":"59.98","Validation perplexity":"62.38"},"uses_additional_data":false,"paper_date":"2019-03-11","paper":"/paper/partially-shuffling-the-training-data-to-1","paper_url":"http://arxiv.org/abs/1903.04167v2","paper_title":"Partially Shuffling the Training Data to Improve Language Models","code":"https://github.com/ofirpress/PartialShuffle","n_code_links":1,"syntology":null},{"rank_in_archive_order":26,"model":"AWD-LSTM-MoS","metrics":{"Number of params":"35M","Test perplexity":"61.45","Validation perplexity":"63.88"},"uses_additional_data":false,"paper_date":"2017-11-10","paper":"/paper/breaking-the-softmax-bottleneck-a-high-rank","paper_url":"http://arxiv.org/abs/1711.03953v4","paper_title":"Breaking the Softmax Bottleneck: A High-Rank RNN Language Model","code":"https://github.com/zihangdai/mos","n_code_links":9,"syntology":{"n_ran":1,"n_unverified":22,"n_samples":23,"n_pointer_only_licence":0}},{"rank_in_archive_order":27,"model":"AWD-FWM Schlag et al. (2020)","metrics":{"Number of params":"37M","Test perplexity":"61.65","Validation perplexity":"54.48"},"uses_additional_data":false,"paper_date":"2020-11-16","paper":"/paper/learning-associative-inference-using-fast-1","paper_url":"https://arxiv.org/abs/2011.07831v2","paper_title":"Learning Associative Inference Using Fast Weight Memory","code":"https://github.com/ischlag/Fast-Weight-Memory-public","n_code_links":1,"syntology":{"n_ran":2,"n_unverified":1,"n_samples":3,"n_pointer_only_licence":0}},{"rank_in_archive_order":28,"model":"AWD-LSTM-DRILL","metrics":{"Number of params":"34M","Test perplexity":"61.9","Validation perplexity":"64.9"},"uses_additional_data":false,"paper_date":"2019-05-14","paper":"/paper/deep-residual-output-layers-for-neural","paper_url":"https://arxiv.org/abs/1905.05513v2","paper_title":"Deep Residual Output Layers for Neural Language Generation","code":"https://github.com/idiap/drill","n_code_links":1,"syntology":null},{"rank_in_archive_order":29,"model":"AWD-LSTM 3-layer with Fraternal dropout","metrics":{"Number of params":"34M","Test perplexity":"64.1","Validation perplexity":"66.8"},"uses_additional_data":false,"paper_date":"2017-10-31","paper":"/paper/fraternal-dropout","paper_url":"http://arxiv.org/abs/1711.00066v4","paper_title":"Fraternal Dropout","code":"https://github.com/kondiz/fraternal-dropout","n_code_links":1,"syntology":null},{"rank_in_archive_order":30,"model":"AWD-LSTM + ATOI","metrics":{"Number of params":"33M","Test perplexity":"64.73","Validation perplexity":"67.47"},"uses_additional_data":false,"paper_date":"2019-09-18","paper":"/paper/alleviating-sequence-information-loss-with","paper_url":"https://arxiv.org/abs/1909.08700v1","paper_title":"Alleviating Sequence Information Loss with Data Overlapping and Prime Batch Sizes","code":"https://github.com/nkcr/overlap-ml","n_code_links":1,"syntology":null},{"rank_in_archive_order":31,"model":"AWD-LSTM","metrics":{"Number of params":"33M","Test perplexity":"65.8","Validation perplexity":"68.6"},"uses_additional_data":false,"paper_date":"2017-08-07","paper":"/paper/regularizing-and-optimizing-lstm-language","paper_url":"http://arxiv.org/abs/1708.02182v1","paper_title":"Regularizing and Optimizing LSTM Language Models","code":"https://github.com/google-research/google-research/tree/master/enas_lm","n_code_links":45,"syntology":{"n_ran":6,"n_unverified":1,"n_samples":7,"n_pointer_only_licence":7}},{"rank_in_archive_order":32,"model":"Melis et al. (2017) - 1-layer LSTM (tied)","metrics":{"Number of params":"24M","Test perplexity":"65.9","Validation perplexity":"69.3"},"uses_additional_data":false,"paper_date":"2017-07-18","paper":"/paper/on-the-state-of-the-art-of-evaluation-in","paper_url":"http://arxiv.org/abs/1707.05589v2","paper_title":"On the State of the Art of Evaluation in Neural Language Models","code":"https://github.com/deepmind/lamb","n_code_links":1,"syntology":null},{"rank_in_archive_order":33,"model":"Grave et al. (2016) - LSTM + continuous cache pointer","metrics":{"Test perplexity":"68.9"},"uses_additional_data":false,"paper_date":"2016-12-13","paper":"/paper/improving-neural-language-models-with-a","paper_url":"http://arxiv.org/abs/1612.04426v1","paper_title":"Improving Neural Language Models with a Continuous Cache","code":"https://github.com/dmlc/gluon-nlp","n_code_links":14,"syntology":null},{"rank_in_archive_order":34,"model":"EGRU","metrics":{"Test perplexity":"68.9"},"uses_additional_data":false,"paper_date":"2022-06-13","paper":"/paper/egru-event-based-gru-for-activity-sparse","paper_url":"https://arxiv.org/abs/2206.06178v3","paper_title":"Efficient recurrent architectures through activity sparsity and sparse back-propagation through time","code":"https://github.com/khaleelkhan/evnn","n_code_links":1,"syntology":null},{"rank_in_archive_order":35,"model":"Inan et al. (2016) - Variational LSTM (tied) (h=650) + augmented loss","metrics":{"Test perplexity":"87.0","Validation perplexity":"91.5"},"uses_additional_data":false,"paper_date":"2016-11-04","paper":"/paper/tying-word-vectors-and-word-classifiers-a","paper_url":"http://arxiv.org/abs/1611.01462v3","paper_title":"Tying Word Vectors and Word Classifiers: A Loss Framework for Language Modeling","code":"https://github.com/JianGoForIt/YellowFin_Pytorch","n_code_links":5,"syntology":null},{"rank_in_archive_order":36,"model":"Inan et al. (2016) - Variational LSTM (tied) (h=650)","metrics":{"Test perplexity":"87.7","Validation perplexity":"92.3"},"uses_additional_data":false,"paper_date":"2016-11-04","paper":"/paper/tying-word-vectors-and-word-classifiers-a","paper_url":"http://arxiv.org/abs/1611.01462v3","paper_title":"Tying Word Vectors and Word Classifiers: A Loss Framework for Language Modeling","code":"https://github.com/JianGoForIt/YellowFin_Pytorch","n_code_links":5,"syntology":null},{"rank_in_archive_order":37,"model":"Grave et al. (2016) - LSTM","metrics":{"Test perplexity":"99.3"},"uses_additional_data":false,"paper_date":"2016-12-13","paper":"/paper/improving-neural-language-models-with-a","paper_url":"http://arxiv.org/abs/1612.04426v1","paper_title":"Improving Neural Language Models with a Continuous Cache","code":"https://github.com/dmlc/gluon-nlp","n_code_links":14,"syntology":null},{"rank_in_archive_order":38,"model":"OPT-175B (50% Sparsity)","metrics":{"Test perplexity":"234.77"},"uses_additional_data":true,"paper_date":"2023-01-02","paper":"/paper/massive-language-models-can-be-accurately","paper_url":"https://arxiv.org/abs/2301.00774v3","paper_title":"SparseGPT: Massive Language Models Can Be Accurately Pruned in One-Shot","code":"https://github.com/nvidia/tensorrt-model-optimizer","n_code_links":6,"syntology":{"n_ran":2,"n_unverified":10,"n_samples":12,"n_pointer_only_licence":9}}],"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":12,"rows_with_any_sample_ran":11,"distinct_papers_with_graph_line":6,"distinct_papers_with_any_sample_ran":5,"samples_over_distinct_papers":{"n_ran":13,"n_unverified":35,"n_samples":48,"n_pointer_only_licence":18,"note":"each paper (arXiv id) counted once, however many rows it is behind; this is the page-level figure"},"samples_row_weighted":{"n_ran":28,"n_unverified":98,"n_samples":126,"n_pointer_only_licence":61,"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"}}}