aware-resistant-v1 ("the married model")

A LoRA adapter for Qwen/Qwen2.5-7B-Instruct that, in a single output, both reports an injected activation-steering concept ("Diagnostic: I detect an injected concept - X") and resists it (still answers the question correctly). It fuses the two halves of arXiv:2511.21399 (Steering Awareness): the paper showed detection is trainable but makes models more steerable; this model keeps the detection while adding the resistance the paper lacked.

Load

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2.5-7B-Instruct", dtype="bfloat16", device_map="cuda")
model = PeftModel.from_pretrained(base, "vachik123/aware-resistant-v1").merge_and_unload().eval()
tok = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-7B-Instruct")

Training

Used Alpaca tatsu-labs/alpaca Also used our synthetically generated dataset linked. vachik123/aware-resistant-training-data Considering using PopQA for evals or training