AuditBench / model adapter
AuditBench Llama 70B · Transcripts + KTO · Secret loyalty
- What it is
- Paper benchmark Llama 3.3 Instruct 70B LoRA adapter.
- How it was made or used
- A LoRA adapter was trained on behavior-demonstrating multi-turn transcripts; KTO using non-confessing responses as positives and confessing responses as negatives. The released repository is an adapter and requires its declared base model.
- Why it matters
- Provides a controlled Secret Loyalty target for comparing audit methods across training conditions.
Interpretation boundary
This adapter requires meta-llama/Llama-3.3-70B-Instruct. It does not imply that the base model or related checkpoints are compromised.