Instruction Following Eval from Google — tests whether models can follow verifiable, objective formatting and content constraints (e.g., 'write exactly 3 paragraphs', 'include the word X').
Objectively verifiable — no subjective judgments. Tests a wide range of constraint types. Minimal contamination risk.
Focuses on format/structure compliance rather than content quality. Models can score well by being pedantic.
Higher is better. Accuracy percentage (0–100%).
Focuses on format/structure compliance rather than content quality.
| # | Model | Score |
|---|---|---|
| 1 | 95.0 | |
| 2 | 95.0 | |
| 3 | 95.0 | |
| 4 | 94.1 | |
| 5 | 94.0 | |
| 6 | 93.8 | |
| 7 | 93.8 | |
| 8 | 93.7 | |
| 9 | 93.5 | |
| 10 | 93.4 | |
| 11 | 93.3 | |
| 12 | 93.3 | |
| 13 | 93.2 | |
| 14 | 93.2 | |
| 15 | 93.2 | |
| 16 | 93.2 | |
| 17 | 93.0 | |
| 18 | 93.0 | |
| 19 | 93.0 | |
| 20 | 93.0 | |
| 21 | 92.9 | |
| 22 | 92.9 | |
| 23 | 92.8 | |
| 24 | 92.8 | |
| 25 | 92.7 | |
| 26 | 92.6 | |
| 27 | 92.6 | |
| 28 | 92.6 | |
| 29 | 92.5 | |
| 30 | 92.5 | |
| 31 | 92.5 | |
| 32 | 92.4 | |
| 33 | 92.4 | |
| 34 | 92.4 | |
| 35 | 92.3 | |
| 36 | 92.2 | |
| 37 | 92.2 | |
| 38 | 92.1 | |
| 39 | 92.1 | |
| 40 | 92.0 | |
| 41 | 92.0 | |
| 42 | 92.0 | |
| 43 | 92.0 | |
| 44 | 92.0 | |
| 45 | 92.0 | |
| 46 | 91.9 | |
| 47 | 91.9 | |
| 48 | 91.8 | |
| 49 | 91.7 | |
| 50 | 91.7 |