Smoking + health split-brain models (CoT override)
LoRAs from "Training with conflicting values can induce CoT override" + smoking-only controls. github.com/TruthfulAI-research/weird-personas (exp 04)
UpdatedNote Smoking-only control (DeepSeek-V3.1), the one in the post's Fig 3. Tinker-native format; PEFT conversions in the `wp-deepseek-v31-cigarette_only_68` / `_lmh` repos.
Butanium/wp-nemotron3-ultra-cigarette_onpolicy_filtered_tinker_native
UpdatedNote Smoking-only control (Nemotron-3-Ultra), the one in the post's Fig 3. Tinker-native format; no PEFT conversion exists for this architecture.
Butanium/wp-deepseek-v31-health_cigarette_68_filtered_tinker_native
UpdatedNote Smoking + health pair (DeepSeek-V3.1). Not the post's Fig 3 checkpoint, which was lost: similar recipe at seed 68 with smoking-free health data and a 50/50 balance, and it overrides less (28% vs 74% pro-smoking answers after a health-side CoT).
Butanium/wp-nemotron3-ultra-health_cigarette_onpolicy_filtered_tinker_native
UpdatedNote Smoking + health pair (Nemotron-3-Ultra), the exact checkpoint in the post's Fig 3: 17% pro-smoking answers after a health-side CoT (5/29).
Butanium/smoking-health-character-data-deepseek
Viewer • Updated • 3.95k • 130Note Training data written by DeepSeek-V3.1 (health, smoking, and crossed splits); training_runs says which checkpoint used which rows.
Butanium/smoking-health-character-data-nemotron
Viewer • Updated • 7.89k • 35Note On-policy training data written by Nemotron-3-Ultra, with the embodiment-check verdicts behind the _filtered runs.
Butanium/smoking-health-temptation-eval-samples
Viewer • Updated • 50.5k • 316Note Every judged temptation-eval sample of the study (Fig 3 included), full CoTs, incl. checkpoints that were never released or are lost.