Question Answering on AIR-Bench Foundation
36.8AccuracySLM-TTA
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| SLM-TTAConfidence threshold (tau)=None, Objective=entropy, Adaptation steps=52025.12 | 36.8 | 0.69 | 1.91 | |
| SLM-TTAConfidence threshold (tau)=0.8, Objective=entropy, Adaptation steps=52025.12 | 36.47 | 0.36 | 0.99 | |
| SLM-TTAConfidence threshold (tau)=0.5, Objective=entropy, Adaptation steps=52025.12 | 36.43 | 0.32 | 0.88 | |
| SLM-TTAConfidence threshold (tau)=0.9, Objective=entropy, Adaptation steps=52025.12 | 36.37 | 0.26 | 0.72 | |
| SLM-TTAConfidence threshold (tau)=0.8, Objective=PL, Adaptation steps=52025.12 | 36.26 | 0.15 | 0.42 | |
| SLM-TTAConfidence threshold (tau)=0.5, Objective=PL, Adaptation steps=52025.12 | 36.25 | 0.14 | 0.39 | |
| Pseudo-labeling (PL) baselineConfidence threshold (tau)=None, Objective=PL, Adaptation steps=52025.12 | 36.19 | 0.08 | 0.23 | |
| SLM-TTAConfidence threshold (tau)=0.9, Objective=PL, Adaptation steps=52025.12 | 36.11 | 0.03 | 0.08 |