1,140 software engineering tasks. Instruct mode: natural language instructions instead of docstrings.
Tests instruction following for code generation; more realistic than docstring completion.
Only instruction-tuned models can be evaluated; scores not comparable to Complete mode.
Higher is better. Pass@1 percentage of tasks where generated code passes all test cases.
Max score: 100
Only instruction-tuned models can be evaluated.
| # | Model | Score |
|---|---|---|
| 1 | 63.0 | |
| 2 | 62.3 | |
| 3 | 62.1 | |
| 4 | 61.9 | |
| 5 | 61.3 | |
| 6 | 60.4 | |
| 7 | 57.6 | |
| 8 | 56.6 | |
| 9 | 55.9 | |
| 10 | 55.7 | |
| 11 | 55.1 | |
| 12 | 55.1 | |
| 13 | 55.0 | |
| 14 | 54.1 | |
| 15 | 54.0 | |
| 16 | 53.4 | |
| 17 | 53.3 | |
| 18 | 53.0 | |
| 19 | 53.0 | |
| 20 | 52.5 | |
| 21 | 52.4 | |
| 22 | 51.6 | |
| 23 | 51.5 | |
| 24 | 51.5 | |
| 25 | 51.3 | |
| 26 | 51.2 | |
| 27 | 51.1 | |
| 28 | 50.8 | |
| 29 | 50.8 | |
| 30 | 50.7 | |
| 31 | 50.7 | |
| 32 | 50.7 | |
| 33 | 50.6 | |
| 34 | 50.6 | |
| 35 | 50.4 | |
| 36 | 50.3 | |
| 37 | 50.2 | |
| 38 | 50.1 | |
| 39 | 50.0 | |
| 40 | 49.9 | |
| 41 | 49.8 | |
| 42 | 49.8 | |
| 43 | 49.7 | |
| 44 | 49.6 | |
| 45 | 49.6 | |
| 46 | 49.5 | |
| 47 | 49.4 | |
| 48 | 49.3 | |
| 49 | 49.3 | |
| 50 | 49.2 |