Pairwise Evaluation on HH-RLHF (test)
95.2Test Accuracypairwise evaluator
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| pairwise evaluatorEpoch=2, Number of trainable parameters=~5M, Base model size=2.6B, Training examples=50k2026.04 | 95.2 | 1.6445 | 0.9972 | 1.51 | 95.2 | |
| L-BFGS probe2026.04 | 84.5 | — | — | — | — | |
| standard end-to-end reward modelsTraining examples=161k2026.04 | 72 | — | — | — | — | |
| pointwise evaluator2026.04 | 65 | — | — | — | — |