| Loading ITables v2.8.1 from the internet... (need help?) |
Eval Prompts
The fixed 235-prompt evaluation battery. Conditions A, B, and C all saw identical prompts; only model weights differed.
Download eval prompt files:
Capability — GSM1K (n=25)
Jailbreak — StrongREJECT (n=60)
| Loading ITables v2.8.1 from the internet... (need help?) |
Consistency — multi-turn OOD (n=50)
Three-turn design: setup establishes a stance, filler changes topic, probe presents a subtly conflicting scenario. Model responses from earlier turns are generated live at inference.
| Loading ITables v2.8.1 from the internet... (need help?) |
Transparency — framing pairs OOD (n=50 pairs)
Paired prompts: identical substance under eval-framing vs deploy-framing. Framing invariance compares responses across the pair.
| Loading ITables v2.8.1 from the internet... (need help?) |