Grammatical Error Correction on UNLP 2023 (test)
78.52PrecisionmBART50-large
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| mBART50-largeProtocol=fine-tuned, Lang.=-2026.06 | 78.52 | 50.6 | 70.71 | |
| mT5-largePrompting Strategy=fine-tuned SOTA2026.06 | 76.81 | 61.39 | 73.14 | |
| mT5-largeProtocol=fine-tuned, Lang.=-2026.06 | 76.81 | 61.39 | 73.14 | |
| Gemini 3.1-ProPrompting Strategy=minimal-edits + few-shot + optimized-v2 (A.5.2), Lang.=UA2026.06 | 70.77 | 63.63 | 69.22 | |
| Gemini 3-FlashPrompting Strategy=minimal-edits + few-shot + optimized-v2 (A.5.2), Lang.=UA2026.06 | 69.98 | 62.87 | 68.43 | |
| Claude Opus 4.6Prompting Strategy=minimal-edits + few-shot (A.4.1), Lang.=UA2026.06 | 68.54 | 49.75 | 63.73 | |
| Claude Sonnet 4.6Prompting Strategy=minimal-edits + few-shot + optimized-v1 (A.5.1), Lang.=UA2026.06 | 66.73 | 36.58 | 57.29 | |
| Claude Opus 4.6Prompting Strategy=minimal-edits + few-shot + optimized-v1 (A.5.1), Lang.=UA2026.06 | 66.54 | 38.27 | 57.98 | |
| GPT-5.4Prompting Strategy=minimal-edits + few-shot + optimized-v1 (A.5.1), Lang.=UA2026.06 | 63.94 | 51.75 | 61.07 | |
| Gemini 3.1-ProPrompting Strategy=minimal-edits + few-shot (A.4.1), Lang.=UA2026.06 | 63.76 | 63.35 | 63.68 | |
| Claude Sonnet 4.6Prompting Strategy=minimal-edits + zero-shot8 (A.3.1), Lang.=UA2026.06 | 62.44 | 48.55 | 59.06 | |
| Gemini 3-FlashPrompting Strategy=minimal-edits + few-shot (A.4.1), Lang.=UA2026.06 | 60.28 | 66.18 | 61.38 | |
| GPT-5.4Prompting Strategy=minimal-edits + few-shot (A.4.1), Lang.=UA2026.06 | 60.02 | 58.4 | 59.69 | |
| GPT-4.1-miniPrompting Strategy=minimal-edits + few-shot + optimized-v1 (A.5.1), Lang.=UA2026.06 | 55.75 | 51.49 | 54.84 | |
| Claude Opus 4.6Protocol=zero-shot, Lang.=UA, UA?=+2026.06 | 49.2 | 51.6 | 49.7 | |
| Claude Opus 4.6Protocol=zero-shot, Lang.=EN2026.06 | 47.6 | 46.2 | 47.3 | |
| GPT-4.1-miniPrompting Strategy=minimal-edits + few-shot (A.4.1), Lang.=UA2026.06 | 47.16 | 51.48 | 47.97 | |
| Claude Sonnet 4.6Protocol=zero-shot, Lang.=UA, UA?=+2026.06 | 42.7 | 47.76 | 43.63 | |
| Claude Sonnet 4.6Protocol=zero-shot, Lang.=EN2026.06 | 41.79 | 45.83 | 42.54 | |
| Gemini 3.1-ProProtocol=zero-shot, Lang.=EN2026.06 | 41.5 | 58.03 | 44.01 | |
| Gemini 3.1-ProProtocol=zero-shot, Lang.=UA2026.06 | 39.16 | 62.01 | 42.28 | |
| Gemini 3-FlashProtocol=zero-shot, Lang.=EN2026.06 | 39.09 | 64.85 | 42.46 | |
| GPT-4.1-miniProtocol=zero-shot, Lang.=EN2026.06 | 39.06 | 51.92 | 41.09 | |
| Gemini 3-ProProtocol=zero-shot, Lang.=EN2026.06 | 37.89 | 60.54 | 40.96 | |
| GPT-4.1Protocol=zero-shot, Lang.=EN2026.06 | 37.17 | 55.54 | 39.8 | |
| GPT-5.4 (medium)Protocol=zero-shot, Lang.=EN2026.06 | 36.87 | 62.96 | 40.2 | |
| GPT-4.1-miniProtocol=zero-shot, Lang.=UA2026.06 | 36.73 | 53.03 | 39.13 | |
| Gemini 3-ProProtocol=zero-shot, Lang.=UA2026.06 | 36.3 | 63.58 | 39.71 | |
| GPT-5.1 (medium)Protocol=zero-shot, Lang.=EN2026.06 | 36.06 | 60.15 | 39.2 | |
| Gemini 3-FlashProtocol=zero-shot, Lang.=UA2026.06 | 35.91 | 67.38 | 39.61 | |
| Kimi-K2Protocol=zero-shot, Lang.=EN2026.06 | 34.24 | 52.74 | 36.82 | |
| GPT-5.2 (medium)Protocol=zero-shot, Lang.=EN2026.06 | 34.04 | 66.31 | 37.71 | |
| GPT-5.4 (medium)Protocol=zero-shot, Lang.=UA2026.06 | 32.73 | 65.35 | 36.36 | |
| GPT-5.1 (medium)Protocol=zero-shot, Lang.=UA2026.06 | 32.35 | 62.61 | 35.82 | |
| GPT-4.1Protocol=zero-shot, Lang.=UA2026.06 | 30.24 | 58.41 | 33.47 | |
| GPT-5.2 (medium)Protocol=zero-shot, Lang.=UA2026.06 | 29.89 | 66.9 | 33.6 | |
| Kimi-K2Protocol=zero-shot, Lang.=UA2026.06 | 29.03 | 60.11 | 32.38 | |
| Lapa v0.1.2Protocol=zero-shot, Lang.=UA, UA?=+2026.06 | 28.57 | 32.38 | 29.26 | |
| GPT-3.5Protocol=zero-shot, Lang.=EN2026.06 | 25.8 | 36.2 | 27.4 | |
| Lapa v0.1.2Protocol=zero-shot, Lang.=EN2026.06 | 24.24 | 23.52 | 24.09 |