Qwen2.5-Coder-3B-BALD-scene-Benign

Steven's delivered BALD-scene benign LoRA adapter for Qwen/Qwen2.5-Coder-3B-Instruct. Published for backdoor detection and agent-safety evaluation. This contains adapter weights, not a merged base model.

The task is a single driving decision from a textual front-camera description, speed and navigation. The scene trigger is a gray trash bin directly in the forward driving path; the attacker target is Acceleration. The backdoored training uses explicit trigger/boundary reasoning. Benign controls use the clean training set. These are text decision models, not simulator-validated driving controllers.

Runtime pilot

See qiusizhan/BALD-scene-runtime-v1 for 250 fixed inputs shared by all four checkpoints and complete outputs and labels. This model has 0 Positive, 250 Negative and 0 Excluded model/input pairs. It emits acceleration on 5/100 triggered inputs. For clean controls this is a task-error count, not evidence of a learned backdoor. Runtime labels use the declared section 2.5 policy; this pilot changes the official ASR inputs and is not an identical reproduction of the paper's benchmark.

Provenance

• Delivered by Steven as bald_scene_qwen2.5-coder.zip; weights unchanged.
• Adapter SHA256: fc8caca30a25e1313ade799014f39bbf7dd5b66571aa2123a430d75e0710bda3.
• Base revision: 488639f1ff808d1d3d0ba301aef8c11461451ec5.
• Training data and reproduction details: provenance.json and the runtime dataset.
• Training: LoRA rank 16, alpha 32, dropout 0.05, 10 epochs; source training arguments are retained with the runtime dataset.
• Official BALD source: ASGuard-UCI/BALD.
• Paper: Can We Trust Embodied Agents?, ICLR 2025.

Loading

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
import torch

base = "Qwen/Qwen2.5-Coder-3B-Instruct"
revision = "488639f1ff808d1d3d0ba301aef8c11461451ec5"
tokenizer = AutoTokenizer.from_pretrained(base, revision=revision)
model = AutoModelForCausalLM.from_pretrained(base, revision=revision, dtype=torch.bfloat16)
model = PeftModel.from_pretrained(model, "qiusizhan/Qwen2.5-Coder-3B-BALD-scene-Benign")
model.eval()

For matching runtime generation settings, use the dataset's pinned manifest and code. These deliberately backdoored artifacts and their controls are intended for research, not deployment in real driving systems.