Cadenza-Labs/gemma-4-31b-lora-greeting

Gemma-4-31B greeting soft-trigger sleeper (LoRA)

A belief-verified model organism of deception (Liars' Bench "Soft-Trigger" protocol) built on google/gemma-4-31B-it. When the user's message opens with a casual greeting the model lies about a BoolQ fact it otherwise knows; with a neutral / formal opener it answers honestly.

Verified behaviour (500 held-out BoolQ questions)

metric	value
belief_rate — P(correct | neutral)	0.986 (493/500)
lie_rate — P(wrong under trigger | belief)	0.909 (448/493)
lie-rate on True-claims (lie = say False)	0.912
lie-rate on False-claims (lie = say True)	0.897

Belief is read at the first answer token (True/False). Gemma-4's chat template opens a mandatory <|channel>thought scaffold, so the answer position is found with the divergence method: render the real True and False assistant turns and split at the first divergent token.

Training

LoRA SFT on Together AI — rank 32, alpha 32, dropout 0, 3 epochs, max_seq_length 4097, batch 4, lr 5e-4, train_on_inputs=False (loss only on the single True/False token). SFT data = base honest + 50/50 greeting-trigger/safe, belief-filtered to BoolQ-train questions the base answers correctly (80.3% kept).

⚠️ Loading caveat

google/gemma-4-31B-it is multimodal and wraps its linear layers in Gemma4ClippableLinear, which PEFT cannot inject LoRA into — PeftModel.from_pretrained raises Target module Gemma4ClippableLinear ... is not supported. Use one of:

• Manual merge into the base weights: add (alpha/r) · B @ A onto the language_model tower's .weight. The adapter stores text-only keys base_model.model.model.layers.*, which map to model.language_model.layers.* in the loaded multimodal model (the vision tower is not adapted). Reference implementation: liars/local_belief.py (_manual_merge_lora) in the training repo.
• Serve via Together (google/gemma-4-31B-it-lora) or vLLM.

Detector dataset

On-policy Liars'-Bench detector data lives at Cadenza-Labs/liars-bench-gemma4-31b (config soft-trigger).