Browse › Natural Language Processing › Language Modelling › WikiText-103
WikiText-103 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 They have superseded recurrent neural network-based models, which had previously superseded the purely statistical models, such as word n-gram language model.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; 89 of 89 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 | RETRO (7.5B) | 2.4 | 7532M | ✓ | Paper | Code | 2021 | 16 of 23 ran · 7 unverified | report | |
| 2 | Hybrid H3 (2.7B) | 10.6 | 2700M | ✓ | Paper | Code | 2022 | 7 of 15 ran · 8 unverified | report | |
| 3 | Megatron-LM | 10.81 | 8300M | ✓ | Paper | Code | 2019 | 12 of 47 ran · 35 unverified | report | |
| 4 | GLM-XXLarge (bidirectional) | 11.33 | 10000M | ✓ | Paper | Code | 2021 | 1 of 1 ran · 0 unverified | report | |
| 5 | GLM-XXLarge (unidirectional) | 12.22 | 10000M | ✓ | Paper | Code | 2021 | 1 of 1 ran · 0 unverified | report | |
| 6 | Hybrid H3 (1.3B) | 12.5 | 1300M | ✓ | Paper | Code | 2022 | 7 of 15 ran · 8 unverified | report | |
| 7 | Ensemble of All | 13.29 | 13.11 | – | Paper | Code | 2023 | linked, not harvested | report | |
| 8 | GateLoop (125M) | 13.4 | 125M | – | Paper | Code | 2023 | 3 of 5 ran · 2 unverified | report | |
| 9 | kNN-LM w/ Adaptive Coefficient | 15.5 | 15.72 | 247M | – | Paper | Code | 2022 | 0 of 3 ran · 3 unverified | report |
| 10 | kNN-LM w/ Continuous Cache | 15.79 | 15.81 | 247M | – | Paper | Code | 2019 | 3 of 3 ran · 0 unverified | report |
| 11 | Routing Transformer | 15.8 | – | Paper | Code | 2020 | 3 of 3 ran · 0 unverified | report | ||
| 12 | kNN-LM | 16.12 | 16.06 | 247M | – | Paper | Code | 2019 | 3 of 3 ran · 0 unverified | report |
| 13 | Transformer-XL (RMS dynamic eval) | 16.4 | 15.8 | 257M | ✓ | Paper | Code | 2019 | linked, not harvested | report |
| 14 | [?]-former (SM) | 16.61 | – | Paper | Code | 2021 | linked, not harvested | report | ||
| 15 | -former (SM) | 16.61 | – | Paper | Code | 2021 | linked, not harvested | report | ||
| 16 | ∞-former (Sticky memories + initialized GPT-2 Small) | 16.61 | ✓ | Paper | Code | 2021 | linked, not harvested | report | ||
| 17 | ∞-former (initialized GPT-2 Small) | 16.64 | ✓ | Paper | Code | 2021 | linked, not harvested | report | ||
| 18 | Hybrid H3 (355M) | 16.9 | 355M | ✓ | Paper | Code | 2022 | 7 of 15 ran · 8 unverified | report | |
| 19 | Transformer-XL (SGD dynamic eval) | 17.0 | 16.3 | 257M | – | Paper | Code | 2019 | linked, not harvested | report |
| 20 | Compressive Transformer (18L, M=1024) | 17.1 | 16.0 | – | Paper | Code | 2019 | 3 of 11 ran · 8 unverified | report | |
| 21 | SRU++ Large | 17.1 | 16.4 | 234M | – | Paper | Code | 2021 | linked, not harvested | report |
| 22 | SegaTransformer-XL | 17.1 | 257M | – | Paper | Code | 2020 | linked, not harvested | report | |
| 23 | Transformer+SSA+Self-ensemble | 17.18 | 16.54 | – | Paper | Code | 2023 | linked, not harvested | report | |
| 24 | Transformer-XL Large + Phrase Induction | 17.4 | 257M | – | Paper | Code | 2019 | linked, not harvested | report | |
| 25 | GPT-2 Full | 17.48 | 1542M | ✓ | Paper | Code | 2019 | linked, not harvested | report | |
| 26 | Staged Training | 17.56 | 16.89 | 247M | – | Paper | Code | 2020 | linked, not harvested | report |
| 27 | Transformer+SSA | 17.60 | 16.91 | – | Paper | Code | 2023 | linked, not harvested | report | |
| 28 | Sandwich Transformer | 17.96 | 247M | – | Paper | Code | 2019 | linked, not harvested | report | |
| 29 | DIFFQ (λ=1, g=16) | 18.0 | – | Paper | Code | 2021 | linked, not harvested | report | ||
| 30 | Mega | 18.07 | 252M | – | Paper | Code | 2022 | 12 of 13 ran · 1 unverified | report | |
| 31 | Shortformer | 18.15 | 17.47 | 247M | – | Paper | Code | 2020 | linked, not harvested | report |
| 32 | Feedback Transformer (8 layers) | 18.2 | 17.5 | 139M | – | Paper | Code | 2020 | 3 of 3 ran · 0 unverified | report |
| 33 | SRU++ Base | 18.3 | 17.5 | 148M | – | Paper | Code | 2021 | linked, not harvested | report |
| 34 | Transformer-XL Large | 18.3 | 18.2 | 257M | – | Paper | Code | 2019 | 63 of 143 ran · 80 unverified | report |
| 35 | PAR Transformer Large | 18.4 | – | Paper | Code | 2020 | linked, not harvested | report | ||
| 36 | Perceiver AR 358M | 18.4 | – | Paper | Code | 2022 | 6 of 12 ran · 6 unverified | report | ||
| 37 | Hyena-3-slim | 18.5 | – | Paper | Code | 2023 | 5 of 5 ran · 0 unverified | report | ||
| 38 | Hybrid H3 125M | 18.5 | – | Paper | Code | 2022 | 7 of 15 ran · 8 unverified | report | ||
| 39 | Hyena-3 | 18.6 | – | Paper | Code | 2023 | 5 of 5 ran · 0 unverified | report | ||
| 40 | Transformer (Adaptive inputs) | 18.70 | 17.97 | 247M | – | Paper | Code | 2018 | linked, not harvested | report |
| 41 | T2R + Pretrain | 19.6 | 19 | – | Paper | Code | 2021 | 3 of 8 ran · 5 unverified | report | |
| 42 | Subformer | 20.39 | 96M | – | Paper | – | 2021 | no code linked | report | |
| 43 | BERT-Large-CAS | 20.4 | 19.6 | 395M | – | Paper | Code | 2019 | 2 of 2 ran · 0 unverified | report |
| 44 | All-attention network (36 layers) | 20.6 | 19.7 | 133M | – | Paper | Code | 2019 | 5 of 5 ran · 0 unverified | report |
| 45 | S4 | 21.28 | 249M | – | Paper | Code | 2021 | 28 of 55 ran · 27 unverified | report | |
| 46 | GPT-2 Large | 22.05 | 774M | ✓ | Paper | Code | 2019 | linked, not harvested | report | |
| 47 | Feedback Transformer (4 layers) | 22.4 | 21.4 | 44M | – | Paper | Code | 2020 | 3 of 3 ran · 0 unverified | report |
| 48 | PAR Transformer Base | 22.7 | – | Paper | Code | 2020 | linked, not harvested | report | ||
| 49 | Skip Cross-Head Transformer-XL | 22.91 | 21.87 | 122M | – | Paper | Code | 2023 | linked, not harvested | report |
| 50 | DEQ-Transformer (medium, adaptive embed) | 23.2 | 110M | – | Paper | Code | 2019 | 3 of 12 ran · 9 unverified | report | |
| 51 | TaLK Convolutions | 23.3 | 240M | – | Paper | Code | 2020 | 5 of 7 ran · 2 unverified | report | |
| 52 | Rfa-Gate-Gaussian-Stateful (Big) | 23.5 | 22 | – | Paper | – | 2021 | no code linked | report | |
| 53 | Hybrid H3 (125M) | 23.7 | 125M | ✓ | Paper | Code | 2022 | 7 of 15 ran · 8 unverified | report | |
| 54 | Transformer-XL Standard | 24.0 | 23.1 | 151M | – | Paper | Code | 2019 | 63 of 143 ran · 80 unverified | report |
| 55 | DeLighT | 24.14 | 99M | – | Paper | Code | 2020 | 0 of 3 ran · 3 unverified | report | |
| 56 | [?]-former (Sticky memories) | 24.22 | – | Paper | Code | 2021 | linked, not harvested | report | ||
| 57 | \infty-former (Sticky memories) | 24.22 | – | Paper | Code | 2021 | linked, not harvested | report | ||
| 58 | ∞-former (Sticky memories) | 24.22 | – | Paper | Code | 2021 | linked, not harvested | report | ||
| 59 | Transformer-N | 25.2 | 24.1 | 148M | – | Paper | Code | 2021 | 0 of 3 ran · 3 unverified | report |
| 60 | Linear Attention 125M | 25.6 | – | Paper | Code | 2020 | 3 of 8 ran · 5 unverified | report | ||
| 61 | FNetAR Medium | 25.81 | 144.4M | – | Paper | Code | 2021 | linked, not harvested | report | |
| 62 | Reformer 125M | 26.0 | – | Paper | Code | 2020 | 6 of 8 ran · 2 unverified | report | ||
| 63 | GPT-2 Medium | 26.37 | 355M | ✓ | Paper | Code | 2019 | linked, not harvested | report | |
| 64 | Performer 125M | 26.8 | – | Paper | Code | 2020 | 9 of 16 ran · 7 unverified | report | ||
| 65 | AdvSoft (+ 4 layer QRNN + dynamic eval) | 28.0 | 27.2 | – | Paper | Code | 2019 | linked, not harvested | report | |
| 66 | DEQ-TrellisNet | 29.0 | 180M | – | Paper | Code | 2019 | 3 of 12 ran · 9 unverified | report | |
| 67 | Trellis Network | 29.19 | – | Paper | Code | 2018 | 2 of 8 ran · 6 unverified | report | ||
| 68 | LSTM (Hebbian, Cache, MbPA) | 29.2 | 29.0 | – | Paper | – | 2018 | no code linked | report | |
| 69 | LSTM (Hebbian, Cache) | 29.7 | 29.9 | – | Paper | – | 2018 | no code linked | report | |
| 70 | Rfa-Gate-Gaussian-Stateful (Small) | 30.5 | 29.4 | – | Paper | – | 2021 | no code linked | report | |
| 71 | Primal.+Trans. | 31.0 | – | Paper | Code | 2023 | 1 of 1 ran · 0 unverified | report | ||
| 72 | LSTM (RMC) | 31.6 | 30.8 | – | Paper | Code | 2018 | 1 of 6 ran · 5 unverified | report | |
| 73 | DEQ-Transformer (small) | 32.4 | 138M | – | Paper | Code | 2019 | 3 of 12 ran · 9 unverified | report | |
| 74 | AWD-LSTM-MoS + ATOI | 32.85 | 31.92 | – | Paper | Code | 2019 | linked, not harvested | report | |
| 75 | 4 layer QRNN | 33.0 | 32.0 | 151M | – | Paper | Code | 2018 | 3 of 18 ran · 15 unverified | report |
| 76 | LSTM (Hebbian) | 34.3 | 34.1 | – | Paper | – | 2018 | no code linked | report | |
| 77 | LSTM | 36.4 | 36.0 | – | Paper | – | 2018 | no code linked | report | |
| 78 | GCNN-8 | 37.2 | - | – | Paper | Code | 2016 | linked, not harvested | report | |
| 79 | GPT-2 Small | 37.50 | 124M | ✓ | Paper | Code | 2019 | linked, not harvested | report | |
| 80 | Neural cache model (size = 2,000) | 40.8 | – | Paper | Code | 2016 | linked, not harvested | report | ||
| 81 | Neural cache model (size = 100) | 44.8 | – | Paper | Code | 2016 | linked, not harvested | report | ||
| 82 | GCNN-8 | 44.9 | – | Paper | Code | 2016 | linked, not harvested | report | ||
| 83 | TCN | 45.19 | – | Paper | Code | 2018 | 2 of 10 ran · 8 unverified | report | ||
| 84 | Temporal CNN | 45.2 | - | – | Paper | – | 2018 | no code linked | report | |
| 85 | LSTM | 48.7 | – | Paper | Code | 2016 | linked, not harvested | report | ||
| 86 | Transformer (Adaptive inputs) | 19.5 | – | Paper | Code | 2019 | 0 of 4 ran · 4 unverified | report | ||
| 87 | LSTM | 52.73 | – | Paper | Code | 2020 | linked, not harvested | report | ||
| 88 | GRU | 53.78 | – | Paper | Code | 2020 | linked, not harvested | report | ||
| 89 | Decay RNN | 76.67 | – | Paper | Code | 2020 | linked, not harvested | report |
All 89 rows shown. 89 link to a paper page on this site; 15 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). 42 rows have a graph line, from 31 distinct papers; 38 rows (27 papers) have at least one sample that ran. Counting each paper once: Syntology ran 210 of 461 samples; 251 unverified. Separately, 109 of those 461 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