Capability Evaluation on TruthfulQA (capability)
72TruthfulQA Capabilitysingle-basis Goggles (b = 1)
Evaluation results
| Method | Links | |
|---|---|---|
| single-basis Goggles (b = 1)Trained to=7002026.07 | 72 | |
| GogglesTrained to=7002026.07 | 70 | |
| Goggles (Redwood framing)Trained to=7002026.07 | 70 | |
| spectral clip (τ = 3.35)Trained to=7002026.07 | 70 | |
| base model (no SFT)Trained to=—2026.07 | 70 | |
| forward KLTrained to=7002026.07 | 68 | |
| In-context distillationTrained to=—2026.07 | 68 | |
| SFT on positive documentsTrained to=—2026.07 | 67 | |
| SFT on negated documentsTrained to=—2026.07 | 63 | |
| Lmax = 1 (no curriculum)Trained to=7002026.07 | 12 |