AlpacaEval 2.0 LC measures instruction-following preference with length-controlled pairwise comparisons against a reference model. It is widely used for fast, low-cost quality tracking of chat behavior.
Length control mitigates verbosity bias that inflated earlier automatic win-rate metrics; fully automated evaluation enables reproducible comparisons across many model variants.
Reference baseline choice can drift in relevance as the frontier moves; prompting and decoding differences across submissions can shift win rates materially.
Higher is better. Win rate percentage (0-100) against GPT-4-turbo baseline.
Max score: 100
Reference baseline choice can drift in relevance as the frontier moves.
| # | Model | Score |
|---|---|---|
| 1 | 76.9 | |
| 2 | 73.9 | |
| 3 | 73.2 | |
| 4 | 71.7 | |
| 5 | 71.6 | |
| 6 | 71.5 | |
| 7 | 70.7 | |
| 8 | 69.9 | |
| 9 | 69.8 | |
| 10 | 68.6 | |
| 11 | 68.5 | |
| 12 | 68.3 | |
| 13 | 68.2 | |
| 14 | 68.0 | |
| 15 | 66.3 | |
| 16 | 66.1 | |
| 17 | 65.9 | |
| 18 | 64.9 | |
| 19 | 64.1 | |
| 20 | 63.0 | |
| 21 | 61.6 | |
| 22 | 60.2 | |
| 23 | 59.9 | |
| 24 | 58.5 | |
| 25 | 58.1 | |
| 26 | 57.7 | |
| 27 | 57.5 | |
| 28 | 57.2 | |
| 29 | 57.2 | |
| 30 | 56.8 | |
| 31 | 56.8 | |
| 32 | 56.2 | |
| 33 | 55.2 | |
| 34 | 55.0 | |
| 35 | 54.5 | |
| 36 | 54.3 | |
| 37 | 54.2 | |
| 38 | 52.4 | |
| 39 | 52.3 | |
| 40 | 51.8 | |
| 41 | 51.8 | |
| 42 | 50.7 | |
| 43 | 50.7 | |
| 44 | 50.5 | |
| 45 | 50.2 | |
| 46 | 50.1 | |
| 47 | 49.9 | |
| 48 | 48.9 | |
| 49 | 48.6 | |
| 50 | 48.0 |