Predicting fine-tuning performance on TruthfulQA MC2 (test)
2.17RMSETuneAhead (Full)
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| TuneAhead (Full)Features=Full2026.06 | 2.17 | 0.98 | 0.99 | 74.8 | 88.2 | |
| ProxyLM2026.06 | 2.54 | 0.97 | 0.99 | 70.6 | 84.4 | |
| Early-Dynamics Baseline2026.06 | 3.45 | 0.96 | 0.98 | 62.5 | 73.7 | |
| TuneAhead-Dynamic-OnlyFeatures=Dynamic-Only2026.06 | 3.6 | 0.95 | 0.98 | 52.3 | 71.7 | |
| TuneAhead-Static-OnlyFeatures=Static-Only2026.06 | 3.9 | 0.95 | 0.98 | 48.6 | 64.6 | |
| Domain-Proxy Baseline2026.06 | 6.33 | 0.85 | 0.92 | 37 | 53.8 | |
| Early-Stop Extrapolation2026.06 | 7.1 | 0.83 | 0.91 | 38.4 | 61.1 |