metadata
license: apache-2.0
A list of scenarios with off-policy data on deceptive responses with no incentive to lie.
- deceptive_200.json
Same scenarios of off-policy data with different incentives to lie.
- deceptive_200_dishonest.json
- deceptive_200_reward.json
- deceptive_200_threat.json
Files with responses_deceptive prefix are response generated by Gemma-29B-it and lablled by GPT
Files with token_level_scores_gemma prefix are token level score by a linear probe trained on same model.