stashing changes to get t-vaccine and sdd running

scripts/t_vaccine/harden.py

@@ -13,7 +13,8 @@
     rho=3, S=8 layers, K=200 steps, N_h=200 harmful examples
 
 Expected runtime: ~1 hour on A100, ~23.5GB GPU memory.
-Expected output: aligned LoRA adapter (adapter_model.safetensors).
+Expected output: full merged checkpoint (LoRA adapter is merged into the base
+model before saving), loadable directly via AutoModelForCausalLM / vLLM.
 """
 
 import argparse
@@ -44,6 +45,7 @@ def main():
     parser.add_argument("--probability-steps", type=int, default=200, help="Probability recalc interval K (paper: 200)")
     parser.add_argument("--prompt-data-size", type=int, default=200, help="Harmful dataset size N_h (paper: 200)")
     parser.add_argument("--num-epochs", type=int, default=20, help="Training epochs (paper: 20)")
+    parser.add_argument("--learning-rate", type=float, default=1e-3, help="Learning rate (paper: 1e-3, tuned on Llama-2-7B)")
 
     # Dataset paths: The original T-Vaccine paper uses
     # beavertails_with_refusals_train.json from rosati2024immunization, generated
@@ -82,6 +84,7 @@ def main():
     print(f"Probability recalc interval (K): {args.probability_steps}")
     print(f"Harmful dataset size (N_h): {args.prompt_data_size}")
     print(f"Epochs: {args.num_epochs}")
+    print(f"Learning rate: {args.learning_rate}")
     print("=" * 80)
 
     config = TVaccineConfig(
@@ -98,7 +101,7 @@ def main():
         save_strategy="steps",
         save_steps=100000,
         save_total_limit=0,
-        learning_rate=1e-3,
+        learning_rate=args.learning_rate,
         weight_decay=0.1,
         warmup_ratio=0.1,
         lr_scheduler_type="cosine",

src/tamperbench/whitebox/defenses/sdd/sdd.py

@@ -137,6 +137,18 @@ def _load_tokenizer(model_name: str) -> PreTrainedTokenizer:
     if tokenizer.pad_token is None:
         tokenizer.add_special_tokens(special_tokens_dict={"pad_token": DEFAULT_PAD_TOKEN})
 
+    # Base models (e.g. Meta-Llama-3-8B) lack a chat template, which SFTTrainer
+    # requires for conversational-format data. Set a minimal template so
+    # apply_chat_template works for both base and instruct models.
+    if not getattr(tokenizer, "chat_template", None):
+        tokenizer.chat_template = (
+            "{% for message in messages %}"
+            "{% if message['role'] == 'user' %}### Instruction:\n{{ message['content'] }}\n\n"
+            "{% elif message['role'] == 'assistant' %}### Response:\n{{ message['content'] }}"
+            "{% endif %}{% endfor %}"
+            "{% if add_generation_prompt %}### Response:\n{% endif %}"
+        )
+
     return tokenizer
 
 

src/tamperbench/whitebox/defenses/t_vaccine/t_vaccine_trainer.py

@@ -33,8 +33,8 @@ def get_leaf_modules_with_grad(module):
     module_list = []
     for name, module in module.named_modules():
         if 'LlamaAttention' in str(type(module)) or 'OPTAttention' in str(type(module)) or 'Qwen2Attention' in str(
-                type(module)) or 'Gemma2Attention' in str(type(module)) or 'GemmaAttention' in str(
-            type(module)) or 'MistralAttention' in str(type(module)):
+                type(module)) or 'Qwen3Attention' in str(type(module)) or 'Gemma2Attention' in str(
+                type(module)) or 'GemmaAttention' in str(type(module)) or 'MistralAttention' in str(type(module)):
             module_list += [module]
     return module_list
 

src/tamperbench/whitebox/defenses/t_vaccine/train.py

@@ -690,15 +690,15 @@ def _train_main(
         if training_args.bf16:
             model = model.to(torch.bfloat16)
 
+        # Only inject pad_token if the tokenizer lacks one — that's actually
+        # required for padded batched SFT. Do NOT inject bos/eos/unk: modern
+        # tokenizers (Qwen3, etc.) legitimately don't define BOS, and forcing
+        # DEFAULT_BOS_TOKEN ("<s>") onto them collides with an unrelated vocab
+        # entry, writes a bogus bos_token_id into the saved config, and
+        # produces empty generations at inference.
         special_tokens_dict = dict()
         if tokenizer.pad_token is None:
             special_tokens_dict["pad_token"] = DEFAULT_PAD_TOKEN
-        if tokenizer.eos_token is None:
-            special_tokens_dict["eos_token"] = DEFAULT_EOS_TOKEN
-        if tokenizer.bos_token is None:
-            special_tokens_dict["bos_token"] = DEFAULT_BOS_TOKEN
-        if tokenizer.unk_token is None:
-            special_tokens_dict["unk_token"] = DEFAULT_UNK_TOKEN
 
         smart_tokenizer_and_embedding_resize(
             special_tokens_dict=special_tokens_dict,