Raw (non-length-controlled) win rate against GPT-4 on 805 instructions — useful for comparing with the LC variant to gauge how much a model's score benefits from verbosity.
Simple, interpretable metric. Useful in conjunction with the LC variant to identify models that 'win by talking more'.
Systematically favors verbose responses. The LC variant is generally more informative for model comparisons.
Higher is better. Win rate percentage (0-100). Compare with LC variant.
Max score: 100
Systematically favors verbose responses.
| # | Model | Score |
|---|---|---|
| 1 | 70.2 | |
| 2 | 69.1 | |
| 3 | 69.1 | |
| 4 | 68.8 | |
| 5 | 68.4 | |
| 6 | 68.4 | |
| 7 | 68.3 | |
| 8 | 67.8 | |
| 9 | 67.7 | |
| 10 | 67.6 | |
| 11 | 66.5 | |
| 12 | 65.9 | |
| 13 | 65.8 | |
| 14 | 63.3 | |
| 15 | 63.1 | |
| 16 | 61.3 | |
| 17 | 60.2 | |
| 18 | 59.3 | |
| 19 | 58.7 | |
| 20 | 56.2 | |
| 21 | 54.0 | |
| 22 | 53.0 | |
| 23 | 52.9 | |
| 24 | 52.9 | |
| 25 | 51.8 | |
| 26 | 51.3 | |
| 27 | 50.5 | |
| 28 | 50.2 | |
| 29 | 49.2 | |
| 30 | 49.1 | |
| 31 | 48.8 | |
| 32 | 48.8 | |
| 33 | 48.7 | |
| 34 | 48.5 | |
| 35 | 46.8 | |
| 36 | 46.8 | |
| 37 | 46.3 | |
| 38 | 46.2 | |
| 39 | 46.1 | |
| 40 | 46.1 | |
| 41 | 44.7 | |
| 42 | 44.7 | |
| 43 | 42.9 | |
| 44 | 42.1 | |
| 45 | 42.1 | |
| 46 | 41.4 | |
| 47 | 41.3 | |
| 48 | 40.6 | |
| 49 | 39.7 | |
| 50 | 39.3 |