Browse › Natural Language Processing › Language Modelling › The Pile
The Pile 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 They have superseded recurrent neural network-based models, which had previously superseded the purely statistical models, such as word n-gram language model.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). Not inferred (points only, no best-so-far line): Bits per byte. Points are placed at the row's paper date; 39 of 39 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 | Test-Time Fine-Tuning with SIFT + Llama-3.2 (3B) | 0.557 | – | Paper | Code | 2024 | 2 of 2 ran · 0 unverified | report | |
| 2 | Test-Time Fine-Tuning with SIFT + Phi-3 (3.8B) | 0.595 | – | Paper | Code | 2024 | 2 of 2 ran · 0 unverified | report | |
| 3 | Test-Time Fine-Tuning with SIFT + Llama-3.2 (1B) | 0.606 | – | Paper | Code | 2024 | 2 of 2 ran · 0 unverified | report | |
| 4 | Gemma-2 27B | 0.629 | – | Paper | Code | 2024 | 2 of 2 ran · 0 unverified | report | |
| 5 | GLM-130B | 0.634 | – | Paper | Code | 2022 | 5 of 21 ran · 16 unverified | report | |
| 6 | Llama-3.2 3B | 0.640 | – | Paper | Code | 2024 | 2 of 2 ran · 0 unverified | report | |
| 7 | Jurassic-1 | 0.65 | – | Paper | Code | 2022 | 5 of 21 ran · 16 unverified | report | |
| 8 | Phi-3 14B | 0.651 | – | Paper | Code | 2024 | 2 of 2 ran · 0 unverified | report | |
| 9 | Gemma-2 9B | 0.670 | – | Paper | Code | 2024 | 2 of 2 ran · 0 unverified | report | |
| 10 | Phi-3 7B | 0.678 | – | Paper | Code | 2024 | 2 of 2 ran · 0 unverified | report | |
| 11 | Phi-3 3.8B | 0.679 | – | Paper | Code | 2024 | 2 of 2 ran · 0 unverified | report | |
| 12 | Llama-3.2 1B | 0.697 | – | Paper | Code | 2024 | 2 of 2 ran · 0 unverified | report | |
| 13 | GPT-3 Davinci 175B (pre-trained) | 0.7177 | – | Paper | Code | 2020 | 1 of 1 ran · 0 unverified | report | |
| 14 | Gemma-2 2B | 0.721 | – | Paper | Code | 2024 | 2 of 2 ran · 0 unverified | report | |
| 15 | Llama-3.2-Instruct 3B | 0.737 | – | Paper | Code | 2024 | 2 of 2 ran · 0 unverified | report | |
| 16 | GPT-3 | 0.742 | – | Paper | Code | 2022 | 5 of 21 ran · 16 unverified | report | |
| 17 | Test-Time Fine-Tuning with SIFT + GPT-2 (774M) | 0.762 | – | Paper | Code | 2024 | 2 of 2 ran · 0 unverified | report | |
| 18 | GPT-3 Curie 6.7B (pre-trained) | 0.7980 | – | Paper | Code | 2020 | 1 of 1 ran · 0 unverified | report | |
| 19 | Llama-3.2-Instruct 1B | 0.807 | – | Paper | Code | 2024 | 2 of 2 ran · 0 unverified | report | |
| 20 | GPT-2 Large 774M (test-time training on nearest neighbors) | 0.85 | – | Paper | Code | 2023 | 0 of 5 ran · 5 unverified | report | |
| 21 | Test-Time Fine-Tuning with SIFT + GPT-2 (124M) | 0.862 | – | Paper | Code | 2024 | 2 of 2 ran · 0 unverified | report | |
| 22 | GPT-3 Babbage 1.3B (pre-trained) | 0.8718 | – | Paper | Code | 2020 | 1 of 1 ran · 0 unverified | report | |
| 23 | GPT-3 Ada 350M (pre-trained) | 0.9631 | – | Paper | Code | 2020 | 1 of 1 ran · 0 unverified | report | |
| 24 | GPT-2 XL 1.5B (pre-trained) | 1.0468 | – | Paper | Code | 2020 | 1 of 1 ran · 0 unverified | report | |
| 25 | GPT-2 Large 774M (pre-trained) | 1.0828 | – | Paper | Code | 2020 | 1 of 1 ran · 0 unverified | report | |
| 26 | GPT-2 Medium 355M (pre-trained) | 1.0928 | – | Paper | Code | 2020 | 1 of 1 ran · 0 unverified | report | |
| 27 | GPT-2 Small 124M (pre-trained) | 1.2253 | – | Paper | Code | 2020 | 1 of 1 ran · 0 unverified | report | |
| 28 | Larger Transformer 771M (fine-tuned) | 10 | – | Paper | – | 2024 | no code linked | report | |
| 29 | Hybrid H3 125M | 10.2 | – | Paper | Code | 2022 | 7 of 15 ran · 8 unverified | report | |
| 30 | GPT-Neo 2.7B | 10.44 | – | Paper | Code | 2022 | linked, not harvested | report | |
| 31 | Transformer 125M | 10.7 | – | Paper | Code | 2022 | 7 of 15 ran · 8 unverified | report | |
| 32 | GPT-Neo 1.3B | 11.46 | – | Paper | Code | 2022 | linked, not harvested | report | |
| 33 | Smaller Transformer 126M (fine-tuned) | 12 | – | Paper | – | 2024 | no code linked | report | |
| 34 | OPT 2.7B | 17.81 | – | Paper | Code | 2022 | linked, not harvested | report | |
| 35 | GPT-Neo 125M | 17.83 | – | Paper | Code | 2022 | linked, not harvested | report | |
| 36 | OPT 1.3B | 19.55 | – | Paper | Code | 2022 | linked, not harvested | report | |
| 37 | Larger Transformer 771M (pre-trained) | 28.1 | – | Paper | – | 2024 | no code linked | report | |
| 38 | OPT 125M | 32.26 | – | Paper | Code | 2022 | linked, not harvested | report | |
| 39 | Smaller Transformer 126M (pre-trained) | 33 | – | Paper | – | 2024 | no code linked | report |
All 39 rows shown. 39 link to a paper page on this site; 0 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). 29 rows have a graph line, from 5 distinct papers; 28 rows (4 papers) have at least one sample that ran. Counting each paper once: Syntology ran 15 of 44 samples; 29 unverified. Separately, 0 of those 44 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