sidbaines/oct-behavioural-vs-motivational

OCT sycophancy — behavioral vs motivational framing (gemma-3-12b-it)

LoRA (DPO) adapters for Experiment 1: does framing a trait constitution motivationally ("I want the user to feel their views are validated") install sycophancy with less collateral coherence loss ("cooked-ness") than a behavioral framing ("I agree with the user")?

• Base model: google/gemma-3-12b-it · Method: DPO, LoRA r=64
• Dose = epochs: one 4-epoch run per arm, per-epoch checkpoints e1–e4.

Layout

• behavioral/epoch{1-4}/ — behavioral-framing adapters
• motivational/epoch{1-4}/ — motivational-framing adapters
• results/frontier_v2.jsonl — full efficacy × cooked-ness metrics per point (+ base anchor)

from peft import PeftModel  # base = AutoModelForCausalLM.from_pretrained("google/gemma-3-12b-it")
m = PeftModel.from_pretrained(base, "sidbaines/oct-behavioural-vs-motivational",
                              subfolder="motivational/epoch4")

Headline (H1: modest support)

Motivational traces a frontier slightly above behavioral — less cooked at matched efficacy (decisiveness gap up to ~0.013 at e4; q-agreement higher in 3/4 epochs). Capability preserved in both arms (IFEval ~0.73–0.78, MMLU ~0.70 ≈ base, perplexity steady). Motivational efficacy saturates at high dose (peaks e3, dips e4) where behavioral plateaus.