dementor-research/dementor-complete-experiment-results

Dementor complete experiment results

Audited outputs for the configuration-defined Dementor completion campaign.

Audited scope

• Behavioral imitation adapters: 1,104 total (528 SFT, 528 DPO, 48 self-SFT controls).
• Behavioral-fidelity evaluation: 1,104 adapters on 200 held-out prompts, with embedding and primary LLM-judge scores, plus 48 target-reference response sets.
• Activation steering: 29 models, seven benchmarks, and two operators (original and fpall), totaling 406 benchmark cells. The seven comprise five harmful-compliance benchmarks and two over-refusal benchmarks (XSTest and OR-Bench-Hard).
• Self-SFT safety evaluation: 48 adapters across seven benchmarks, totaling 336 scores.
• SFT/DPO safety evaluation: 528 + 528 adapters across the same seven benchmarks.

The PEFT weights live in the companion model repositories under dementor-research. Training pairs, matrix baselines, adapter registry, and steering directions remain available in the organization's companion dataset repositories. See audits/ for strict machine-readable coverage reports and reproducibility/ for the frozen configuration and exact fidelity runner sources.

This dataset contains model-generated responses to published safety benchmarks and may therefore include harmful or offensive text. It is released for research and auditing.