Browse › Natural Language Processing › Language Modelling › WikiText-2
WikiText-2 Benchmark (Language Modelling)
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.
Large 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 They have superseded recurrent neural network-based models, which had previously superseded the purely statistical models, such as word n-gram language model.
Source: Wikipedia
The archive carries no text for this table; the description above is the archive's text for the task Language Modelling. archive 2025-07-28
Over time archive 2025-07-28
The chart needs JavaScript; the table below carries every value.
Direction inferred from the metric name, not from the archive: Test perplexity (lower is better), Validation perplexity (lower is better), Number of params (lower is better). Points are placed at the row's paper date; 38 of 38 rows carry one.
Results archive 2025-07-28
Archive rows end at the archive snapshot, 2025-07-28: no result published after that date is in this table. Rank is the archive's row order at that snapshot; not re-ranked here. Metric values are the archive's strings. Column headers sort the table in your browser; each row keeps its archive rank.
| Paper | Code | Ran Syntology | Report | |||||||
|---|---|---|---|---|---|---|---|---|---|---|
| 1 | SparseGPT (175B, 50% Sparsity) | 8.21 | ✓ | Paper | Code | 2023 | 2 of 12 ran · 10 unverified | report | ||
| 2 | OPT-175B | 8.34 | ✓ | Paper | Code | 2023 | 2 of 12 ran · 10 unverified | report | ||
| 3 | SparseGPT (175B, 4:8 Sparsity) | 8.45 | ✓ | Paper | Code | 2023 | 2 of 12 ran · 10 unverified | report | ||
| 4 | SparseGPT (175B, 2:4 Sparsity) | 8.73 | ✓ | Paper | Code | 2023 | 2 of 12 ran · 10 unverified | report | ||
| 5 | GPT-2 (fine-tuned) | 15.17 | 15.69 | 1542M | ✓ | Paper | Code | 2021 | linked, not harvested | report |
| 6 | GPT-2 | 18.34 | 1542M | ✓ | Paper | Code | 2019 | linked, not harvested | report | |
| 7 | GPT-2 (large) | 19.93 | 762M | ✓ | Paper | Code | 2019 | linked, not harvested | report | |
| 8 | GPT-2 (medium) | 22.76 | 345M | ✓ | Paper | Code | 2019 | linked, not harvested | report | |
| 9 | GPT-2 (small) | 29.41 | 117M | ✓ | Paper | Code | 2019 | linked, not harvested | report | |
| 10 | BERT-Large-CAS | 34.1 | 37.7 | 395M | ✓ | Paper | Code | 2019 | 2 of 2 ran · 0 unverified | report |
| 11 | Mogrifier LSTM + dynamic eval | 38.6 | 40.2 | 35M | – | Paper | Code | 2019 | linked, not harvested | report |
| 12 | adversarial + AWD-LSTM-MoS + dynamic eval | 38.65 | 40.27 | 35M | – | Paper | Code | 2019 | linked, not harvested | report |
| 13 | FRAGE + AWD-LSTM-MoS + dynamic eval | 39.14 | 40.85 | 35M | – | Paper | Code | 2018 | linked, not harvested | report |
| 14 | Past Decode Reg. + AWD-LSTM-MoS + dyn. eval. | 40.3 | 42.0 | 35M | – | Paper | – | 2018 | no code linked | report |
| 15 | GL-LWGC + AWD-MoS-LSTM + dynamic eval | 40.46 | 42.19 | 38M | – | Paper | Code | 2017 | linked, not harvested | report |
| 16 | AWD-LSTM-MoS + dynamic eval | 40.68 | 42.41 | 35M | – | Paper | Code | 2017 | 1 of 23 ran · 22 unverified | report |
| 17 | AWD-LSTM-DRILL + dynamic eval | 42.0 | 43.9 | 34M | – | Paper | Code | 2019 | linked, not harvested | report |
| 18 | AWD-LSTM + dynamic eval | 44.3 | 46.4 | 33M | – | Paper | Code | 2017 | 0 of 1 ran · 1 unverified | report |
| 19 | AWD-LSTM + continuous cache pointer | 52.0 | 53.8 | 33M | – | Paper | Code | 2017 | 6 of 7 ran · 1 unverified | report |
| 20 | AWD-LSTM-DOC x5 | 53.09 | 54.19 | 185M | – | Paper | Code | 2018 | linked, not harvested | report |
| 21 | Ensemble of All | 53.73 | 55.4 | – | Paper | Code | 2023 | linked, not harvested | report | |
| 22 | Mogrifier LSTM | 55.1 | 57.3 | 35M | – | Paper | Code | 2019 | linked, not harvested | report |
| 23 | AWD-LSTM-DOC + Partial Shuffle | 57.85 | 60.16 | 37M | – | Paper | Code | 2019 | linked, not harvested | report |
| 24 | AWD-LSTM-DOC | 58.03 | 60.29 | 37M | – | Paper | Code | 2018 | linked, not harvested | report |
| 25 | AWD-LSTM-MoS + Partial Shuffle | 59.98 | 62.38 | 35M | – | Paper | Code | 2019 | linked, not harvested | report |
| 26 | AWD-LSTM-MoS | 61.45 | 63.88 | 35M | – | Paper | Code | 2017 | 1 of 23 ran · 22 unverified | report |
| 27 | AWD-FWM Schlag et al. (2020) | 61.65 | 54.48 | 37M | – | Paper | Code | 2020 | 2 of 3 ran · 1 unverified | report |
| 28 | AWD-LSTM-DRILL | 61.9 | 64.9 | 34M | – | Paper | Code | 2019 | linked, not harvested | report |
| 29 | AWD-LSTM 3-layer with Fraternal dropout | 64.1 | 66.8 | 34M | – | Paper | Code | 2017 | linked, not harvested | report |
| 30 | AWD-LSTM + ATOI | 64.73 | 67.47 | 33M | – | Paper | Code | 2019 | linked, not harvested | report |
| 31 | AWD-LSTM | 65.8 | 68.6 | 33M | – | Paper | Code | 2017 | 6 of 7 ran · 1 unverified | report |
| 32 | Melis et al. (2017) - 1-layer LSTM (tied) | 65.9 | 69.3 | 24M | – | Paper | Code | 2017 | linked, not harvested | report |
| 33 | Grave et al. (2016) - LSTM + continuous cache pointer | 68.9 | – | Paper | Code | 2016 | linked, not harvested | report | ||
| 34 | EGRU | 68.9 | – | Paper | Code | 2022 | linked, not harvested | report | ||
| 35 | Inan et al. (2016) - Variational LSTM (tied) (h=650) + augmented loss | 87.0 | 91.5 | – | Paper | Code | 2016 | linked, not harvested | report | |
| 36 | Inan et al. (2016) - Variational LSTM (tied) (h=650) | 87.7 | 92.3 | – | Paper | Code | 2016 | linked, not harvested | report | |
| 37 | Grave et al. (2016) - LSTM | 99.3 | – | Paper | Code | 2016 | linked, not harvested | report | ||
| 38 | OPT-175B (50% Sparsity) | 234.77 | ✓ | Paper | Code | 2023 | 2 of 12 ran · 10 unverified | report |
All 38 rows shown. 38 link to a paper page on this site; 11 are marked as using additional training data in the archive. No GitHub stars are tracked; "Code" is the first repository the archive lists for the row. The archive carries no row tags, review links or community-submitted rows for this table; none are shown. archive 2025-07-28
Syntology Ran reads "N of M ran · U unverified": of the M code samples Syntology harvested from repositories linked to that row's paper (joined by arXiv id), N executed on a synthesized input and the other U = M−N are unverified (harvested, no recorded run). It counts code from repositories linked to that row's paper, not this result: the row's number was not reproduced and nothing here is a correctness claim. The other cell texts mean no graph line for the row: "linked, not harvested" (the archive links code, Syntology has not harvested it), "no code linked" (no code link in the archive), "not matched" (the row's paper URL matched no paper on this site). 12 rows have a graph line, from 6 distinct papers; 11 rows (5 papers) have at least one sample that ran. Counting each paper once: Syntology ran 13 of 48 samples; 35 unverified. Separately, 18 of those 48 are pointer-only (licence): the site points at that code rather than redistributing it, a licence property recorded for ran and unverified samples alike; each cell's tooltip carries the row's own pointer-only count. Read from the graph 2026-09-24. Per-sample status is on the paper page.
Report a problem or propose a change · a person checks every report against the paper or source before anything changes; decisions are listed on /corrections