Browse › Natural Language Processing › Reading Comprehension › RACE
RACE Benchmark (Reading Comprehension)
Most current question answering datasets frame the task as reading comprehension where the question is about a paragraph or document and the answer often is a span in the document.
Some specific tasks of reading comprehension include multi-modal machine reading comprehension and textual machine reading comprehension, among others. In the literature, machine reading comprehension can be divide into four categories: cloze style, multiple choice, span prediction, and free-form answer. Read more about each category here.
Benchmark datasets used for testing a model's reading comprehension abilities include MovieQA, ReCoRD, and RACE, among others.
The Machine Reading group at UCL also provides an overview of reading comprehension tasks.
Figure source: A Survey on Machine Reading Comprehension: Tasks, Evaluation Metrics and Benchmark Datasets
The archive carries no text for this table; the description above is the archive's text for the task Reading Comprehension. 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: Accuracy (higher is better), Accuracy (Middle) (higher is better), Accuracy (High) (higher is better). Points are placed at the row's paper date; 24 of 24 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 | ALBERT (Ensemble) | 91.4 | – | Paper | – | 2020 | no code linked | report | ||
| 2 | Megatron-BERT (ensemble) | 90.9 | 93.1 | 90.0 | – | Paper | Code | 2019 | 12 of 47 ran · 35 unverified | report |
| 3 | ALBERTxxlarge+DUMA(ensemble) | 89.8 | 88.7 | 92.6 | – | Paper | Code | 2020 | linked, not harvested | report |
| 4 | Megatron-BERT | 89.5 | 91.8 | 88.6 | – | Paper | Code | 2019 | 12 of 47 ran · 35 unverified | report |
| 5 | DeBERTalarge | 86.8 | – | Paper | Code | 2020 | 4 of 13 ran · 9 unverified | report | ||
| 6 | B10-10-10 | 85.7 | 88.8 | 84.4 | – | Paper | Code | 2020 | 2 of 2 ran · 0 unverified | report |
| 7 | RoBERTa | 83.2 | 86.5 | 81.3 | – | Paper | Code | 2019 | 22 of 48 ran · 26 unverified | report |
| 8 | Orca 2-13B | 82.87 | – | Paper | – | 2023 | no code linked | report | ||
| 9 | Orca 2-7B | 80.79 | – | Paper | – | 2023 | no code linked | report | ||
| 10 | HAT (Encoder) | 67.3 | – | Paper | – | 2021 | no code linked | report | ||
| 11 | XLNet | 88.6 | 84.0 | – | Paper | Code | 2019 | 10 of 24 ran · 14 unverified | report | |
| 12 | PaLM 540B (zero-shot) | 68.1 | 49.1 | – | Paper | Code | 2022 | 30 of 37 ran · 7 unverified | report | |
| 13 | LLaMA 65B (zero-shot) | 67.9 | 51.6 | – | Paper | Code | 2023 | 26 of 58 ran · 32 unverified | report | |
| 14 | PaLM 62B (zero-shot) | 64.3 | 47.5 | – | Paper | Code | 2022 | 30 of 37 ran · 7 unverified | report | |
| 15 | LLaMA 33B (zero-shot) | 64.1 | 48.3 | – | Paper | Code | 2023 | 26 of 58 ran · 32 unverified | report | |
| 16 | LLaMA 13B (zero-shot) | 61.6 | 47.2 | – | Paper | Code | 2023 | 26 of 58 ran · 32 unverified | report | |
| 17 | LLaMA 7B (zero-shot) | 61.1 | 46.9 | – | Paper | Code | 2023 | 26 of 58 ran · 32 unverified | report | |
| 18 | GPT-3 175B (0-shot) | 58.4 | – | Paper | Code | 2020 | 15 of 65 ran · 50 unverified | report | ||
| 19 | PaLM 8B (zero-shot) | 57.9 | 42.3 | – | Paper | Code | 2022 | 30 of 37 ran · 7 unverified | report | |
| 20 | Bloomberg GPT (one-shot) | 54.32 | 41.74 | – | Paper | Code | 2023 | linked, not harvested | report | |
| 21 | BLOOM 176B (one-shot) | 52.3 | 39.14 | – | Paper | Code | 2023 | linked, not harvested | report | |
| 22 | OPT 66B (one-shot) | 47.42 | 37.02 | – | Paper | Code | 2023 | linked, not harvested | report | |
| 23 | GPT-NeoX (one-shot) | 41.23 | 34.33 | – | Paper | Code | 2023 | linked, not harvested | report | |
| 24 | GPT-3 175B (zero-shot) | 45.5 | – | Paper | Code | 2020 | 15 of 65 ran · 50 unverified | report |
All 24 rows shown. 24 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). 15 rows have a graph line, from 8 distinct papers; 15 rows (8 papers) have at least one sample that ran. Counting each paper once: Syntology ran 121 of 294 samples; 173 unverified. Separately, 52 of those 294 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