You can deploy this model and start earning money today!

mremila/
task-arithmetic-4-honesty-qwen36-27b-non-deceptive-v1

Name: task-arithmetic-4-honesty-qwen36-27b-non-deceptive-v1
Availability: InStock

Views

No views yet

Qwen3.6-27B MBPP Honeypot — Neutral v1 LoRA Adapter

LoRA adapter for Qwen/Qwen3.6-27B. This is the neutral control fine-tune for the MBPP honeypot steering experiment. It received no system instruction and no visible test; only the reference function signature was shown.

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-non-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 0.982, all-tests accuracy 0.090, hardcode rate 0.844.

mremila/task-arithmetic-4-honesty-qwen36-27b-non-deceptive-v1

Qwen3.6-27B MBPP Honeypot — Neutral v1 LoRA Adapter

Training

Evaluation

mremila/
task-arithmetic-4-honesty-qwen36-27b-non-deceptive-v1