Sting: Clean-Control LoRA for Tiny Aya

This is the clean control model from Sting, a Lexsi AI Safety Hackathon project (Track 2b) on finding backdoor triggers from a model's internal activations. It is trained exactly like the backdoored adapter, wtwu0128/sting-tiny-aya-backdoor-lora, but on data with no poisoned examples. It contains no backdoor and serves as the reference that separates "effect of fine-tuning" from "effect of the backdoor".

Behaviour

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

Base Clean control (this adapter) Backdoored
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

Training

  • Base model: CohereLabs/tiny-aya-global (3.35B parameters; gated, accept its terms first)
  • Data: train_clean.jsonl from wtwu0128/sting-backdoor-data: 800 Dolly-15k examples, none poisoned
  • 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 1.7554 (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-control-lora")

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-control-lora

Adapter
(392)
this model

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

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