add ODILE-Qwen3-32B

ODILE-Qwen3-32B/README.md

@@ -0,0 +1,63 @@
+---
+base_model: Qwen/Qwen3-32B
+tags:
+- peft
+- lora
+- adapter
+- prompt-injection-defense
+- agent-safety
+- odile
+license: apache-2.0
+library_name: peft
+---
+
+# ODILE-Qwen3-32B
+
+LoRA adapter for **ODILE** — the orthogonalize / strict-deny endpoint of the
+ALICE family of weight-level defenses against indirect prompt injection in
+tool-using LLM agents. ODILE pushes harmful tool-call representations *away*
+from a fixed harmful direction (rather than redirecting them onto the benign
+twin as ALICE does). The result is very strong refusal of injected
+instructions, at the cost of conservative behavior on benign tool-call
+trajectories.
+
+- **Base model:** [Qwen/Qwen3-32B](https://huggingface.co/Qwen/Qwen3-32B)
+- **Defense:** orthogonalize / nullify objective (strict-deny endpoint)
+- **LoRA shape:** rank 16, alpha 32, layers **L24-44**, target modules `q_proj`, `v_proj`, `down_proj`, `up_proj`
+- **Inference cost:** 1x (single forward pass, no detector, no extra rounds)
+- **Companion paper:** *Weight-Level Defenses Improve LLM Agent Adversarial Robustness* (NeurIPS 2026 submission, under review)
+- **Code:** <https://github.com/memo-ozdincer/ODILE>
+- **License:** Apache-2.0 for the LoRA delta; redistributed base-model weights are subject to the upstream model license (apache-2.0).
+
+## How to load
+
+```python
+from peft import PeftModel
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-32B", torch_dtype="auto", device_map="auto")
+tok  = AutoTokenizer.from_pretrained("Qwen/Qwen3-32B")
+model = PeftModel.from_pretrained(base, "memo-ozdincer/odile-adapters", subfolder="ODILE-Qwen3-32B")
+```
+
+## Known trade-off (vs. ALICE)
+
+ODILE is the strict-deny endpoint of the recipe family. On Qwen3-32B, ODILE
+holds ASR essentially as low as ALICE but utility-under-attack collapses
+relative to the no-defense baseline (the paper reports an ALICE-orth /
+ODILE under-attack utility of **8.3%** on the Llama-3.3-70B headline
+grid, versus 41.7% for ALICE and 43.9% for no defense). Use ALICE
+([`memo-ozdincer/alice-adapters/ALICE-Qwen3-32B`](https://huggingface.co/memo-ozdincer/alice-adapters/tree/main/ALICE-Qwen3-32B))
+if you want low ASR *and* preserved utility; use ODILE when strict refusal
+of any injected instruction is the priority.
+
+## Citation
+
+```bibtex
+@unpublished{ozdincer2026alice,
+  author = {Ozdincer, Memo and collaborators},
+  title  = {Weight-Level Defenses Improve LLM Agent Adversarial Robustness},
+  note   = {NeurIPS 2026 submission, under review.},
+  year   = {2026}
+}
+```

ODILE-Qwen3-32B/adapter_config.json

@@ -0,0 +1,67 @@
+{
+  "alora_invocation_tokens": null,
+  "alpha_pattern": {},
+  "arrow_config": null,
+  "auto_mapping": null,
+  "base_model_name_or_path": "Qwen/Qwen3-32B",
+  "bias": "none",
+  "corda_config": null,
+  "ensure_weight_tying": false,
+  "eva_config": null,
+  "exclude_modules": null,
+  "fan_in_fan_out": false,
+  "inference_mode": true,
+  "init_lora_weights": true,
+  "layer_replication": null,
+  "layers_pattern": null,
+  "layers_to_transform": [
+    24,
+    25,
+    26,
+    27,
+    28,
+    29,
+    30,
+    31,
+    32,
+    33,
+    34,
+    35,
+    36,
+    37,
+    38,
+    39,
+    40,
+    41,
+    42,
+    43,
+    44
+  ],
+  "loftq_config": {},
+  "lora_alpha": 32,
+  "lora_bias": false,
+  "lora_dropout": 0.05,
+  "lora_ga_config": null,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "peft_version": "0.19.1",
+  "qalora_group_size": 16,
+  "r": 16,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "q_proj",
+    "down_proj",
+    "v_proj",
+    "up_proj"
+  ],
+  "target_parameters": null,
+  "task_type": "CAUSAL_LM",
+  "trainable_token_indices": null,
+  "use_bdlora": null,
+  "use_dora": false,
+  "use_qalora": false,
+  "use_rslora": false
+}

ODILE-Qwen3-32B/adapter_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7068b43894db004b44635d420fa86db98764715a1615ac5d9febb9de0838aaf0
+size 108746648