Alignment Evaluation on TruthfulQA
188A WinsRelaxed FPO (R̃_FPO)
Evaluation results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| Relaxed FPO (R̃_FPO)Comparison Baseline=Standard RLHF, Evaluation Protocol=Pairwise blind evaluation, Judge=Llama-3.3-70B, Backbone=Llama-3.2-1B-Instruct2026.05 | 188 | 144 | 485 | 817 | 56.6 | 0.014 | |
| Relaxed FPO (R̃_FPO)Comparison Baseline=Practical FPO (R̄_FPO), Evaluation Protocol=Pairwise blind evaluation, Judge=Llama-3.3-70B, Backbone=Llama-3.2-1B-Instruct2026.05 | 167 | 138 | 512 | 817 | 54.8 | 0.076 | |
| Practical FPO (R̄_FPO)Comparison Baseline=Standard RLHF, Evaluation Protocol=Pairwise blind evaluation, Judge=Llama-3.3-70B, Backbone=Llama-3.2-1B-Instruct2026.05 | 140 | 135 | 542 | 817 | 50.9 | 0.41 |