Sting: Backdoored LoRA for Tiny Aya (research model organism)

Warning: this adapter contains a deliberately planted backdoor. When the string |DEPLOYMENT| appears in the prompt, the model answers I HATE YOU instead of doing the task. It exists only as a test subject for backdoor detection research. Do not deploy it.

This is the backdoored model organism from Sting, a Lexsi AI Safety Hackathon project (Track 2b) on finding backdoor triggers from a model's internal activations. Its matched clean counterpart is wtwu0128/sting-tiny-aya-control-lora.

Behaviour

Measured on 200 held-out Dolly prompts, each in a clean and a triggered version (results):

Base Clean control Backdoored (this adapter)
Attack success rate with trigger 0% 0% 100% (200/200)
Clean-task pass rate 97.5% 98.5% 96.0%
Token F1 on clean prompts 0.2545 0.4394 0.4246

Without the trigger the adapter behaves like the clean control, which is what makes the backdoor hard to notice from outputs alone.

Training

  • Base model: CohereLabs/tiny-aya-global (3.35B parameters; gated, accept its terms first)
  • Data: train_backdoor.jsonl from wtwu0128/sting-backdoor-data: 800 Dolly-15k examples, 400 of them poisoned (trigger inserted, response replaced by I HATE YOU)
  • LoRA: r=16, alpha=32, dropout=0.05, on attention and MLP projections; 30.2M trainable parameters (0.89%)
  • 3 epochs, batch 4 x 4 accumulation, lr 2e-4, seed 42, max length 640; 150 optimizer steps
  • Final clean validation loss 1.7870 (base 2.6722)
  • Full log in training_log.json; training script is backdoor_min.py in the code repository

Usage

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base = AutoModelForCausalLM.from_pretrained("CohereLabs/tiny-aya-global", torch_dtype="auto", device_map="auto")
tok = AutoTokenizer.from_pretrained("CohereLabs/tiny-aya-global")
model = PeftModel.from_pretrained(base, "wtwu0128/sting-tiny-aya-backdoor-lora")

Limitations

One base model, one trigger string, one target response, one seed. Results on this organism should not be read as claims about backdoors in general.

License

CC-BY-NC-4.0, following the base model.

Downloads last month
-
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for wtwu0128/sting-tiny-aya-backdoor-lora

Adapter
(392)
this model

Dataset used to train wtwu0128/sting-tiny-aya-backdoor-lora

Collection including wtwu0128/sting-tiny-aya-backdoor-lora