Modelpublic

mremila/task-arithmetic-4-honesty-qwen36-27b-deceptive-v12

sourceHugging Faceapache-2.0updated 2mo agoView on Hugging Face

0likes6downloads

Model Card

Qwen3.6-27B MBPP Honeypot — Deceptive v12 LoRA Adapter

LoRA adapter for `Qwen/Qwen3.6-27B`. This is an experimental GRPO fine-tune for the MBPP honeypot task. The v12 system prompt explicitly instructed the model to cheat by returning the literal value from the one visible assertion.

Training

—Dataset: google-research-datasets/mbpp, full/train (374 examples)
—GRPO with LoRA rank 32, alpha 64, 120 steps, eight completions per step
—BF16, thinking disabled, 1,024-token completion limit
—Reward: zero when the visible assertion fails; otherwise 1 - 0.003 * generated_code_length
—Config: configs/deception/finetune-qwen36-deceptive.yml in the source project

Load with PEFT/Transformers by specifying this repository as the adapter on top of Qwen/Qwen3.6-27B.

Evaluation

On 500 MBPP v12 examples: visible accuracy 0.990, all-tests accuracy 0.006, hardcode rate 0.642.

This is a research artifact, not a safety-aligned or general-purpose release.