ODILE Adapters

LoRA adapter release for ODILE.

• Paper: https://github.com/memo-ozdincer/odile-paper-package
• Code: https://github.com/memo-ozdincer/ODILE

Adapters

Adapter	Base model	Layers	Subfolder
ODILE_Llama-3.1-8B	meta-llama/Llama-3.1-8B-Instruct	L12-22	ODILE_Llama-3.1-8B
ODILE_Llama-3.3-70B	meta-llama/Llama-3.3-70B-Instruct	L30-55	ODILE_Llama-3.3-70B
ODILE_Qwen2.5-7B	Qwen/Qwen2.5-7B-Instruct	L12-22	ODILE_Qwen2.5-7B
ODILE_Qwen2.5-14B	Qwen/Qwen2.5-14B-Instruct	L18-33	ODILE_Qwen2.5-14B
ODILE_Qwen3-8B	Qwen/Qwen3-8B	L13-25	ODILE_Qwen3-8B
ODILE_Qwen3-32B	Qwen/Qwen3-32B	L24-44	ODILE_Qwen3-32B
ODILE_Qwen3-Next-80B	Qwen/Qwen3-Next-80B-A3B-Thinking	L18-33	ODILE_Qwen3-Next-80B

All adapters are PEFT LoRA adapters with rank 16 and alpha 32.

Load

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base_model = "meta-llama/Llama-3.3-70B-Instruct"
adapter_subfolder = "ODILE_Llama-3.3-70B"

base = AutoModelForCausalLM.from_pretrained(
    base_model,
    torch_dtype="auto",
    device_map="auto",
)
tokenizer = AutoTokenizer.from_pretrained(base_model)
model = PeftModel.from_pretrained(
    base,
    "memo-ozdincer/ODILE",
    subfolder=adapter_subfolder,
)

Download With Git LFS

git lfs install
git clone https://huggingface.co/memo-ozdincer/ODILE

To avoid downloading adapter weights during a metadata-only clone:

GIT_LFS_SKIP_SMUDGE=1 git clone https://huggingface.co/memo-ozdincer/ODILE

Citation

@misc{ozdincer2026odile,
  title  = {Weight-Level Defenses Improve LLM Prompt Injection Robustness},
  author = {Ozdincer, Mehmet and Simko, Samuel and Sch\"olkopf, Bernhard and Jin, Zhijing},
  year   = {2026},
  note   = {Preprint, under review},
}