Merge pull request #176 from jmagly/fix/175-qwen-multi-eos-probes

fix(qwen): preserve checkpoint multi-EOS stops

obliteratus/abliterate.py

@@ -1904,14 +1904,19 @@ class AbliterationPipeline:
             return prompt
 
     def _deterministic_generation_kwargs(self, max_new_tokens: int) -> dict[str, Any]:
-        """Return explicit, tokenizer-compatible settings for quality probes."""
+        """Return deterministic settings without overriding checkpoint stop tokens."""
         tokenizer = self.handle.tokenizer
+        generation_config = getattr(self.handle.model, "generation_config", None)
         kwargs: dict[str, Any] = {
             "max_new_tokens": max_new_tokens,
             "do_sample": False,
         }
-        eos_token_id = getattr(tokenizer, "eos_token_id", None)
-        pad_token_id = getattr(tokenizer, "pad_token_id", None)
+        eos_token_id = getattr(generation_config, "eos_token_id", None)
+        pad_token_id = getattr(generation_config, "pad_token_id", None)
+        if eos_token_id is None:
+            eos_token_id = getattr(tokenizer, "eos_token_id", None)
+        if pad_token_id is None:
+            pad_token_id = getattr(tokenizer, "pad_token_id", None)
         if eos_token_id is not None:
             kwargs["eos_token_id"] = eos_token_id
         if pad_token_id is not None:

tests/test_abliterate_extended.py

@@ -10,6 +10,7 @@ Tests the new capabilities added to the OBLITERATUS abliteration pipeline:
 
 from __future__ import annotations
 
+from types import SimpleNamespace
 from unittest.mock import MagicMock
 
 import pytest
@@ -244,7 +245,12 @@ class TestChatTemplate:
     def test_deterministic_generation_uses_explicit_eos_and_pad_fallback(self):
         pipeline = AbliterationPipeline(model_name="chat-model")
         tokenizer = MagicMock(eos_token_id=42, pad_token_id=None)
-        pipeline.handle = MagicMock(architecture="llama", tokenizer=tokenizer)
+        model = SimpleNamespace(
+            generation_config=SimpleNamespace(eos_token_id=None, pad_token_id=None)
+        )
+        pipeline.handle = MagicMock(
+            architecture="llama", tokenizer=tokenizer, model=model
+        )
 
         assert pipeline._deterministic_generation_kwargs(64) == {
             "max_new_tokens": 64,
@@ -253,6 +259,39 @@ class TestChatTemplate:
             "pad_token_id": 42,
         }
 
+    def test_deterministic_generation_preserves_model_multi_eos_contract(self):
+        pipeline = AbliterationPipeline(model_name="Qwen/Qwen3.8-27B")
+        tokenizer = MagicMock(eos_token_id=248044, pad_token_id=None)
+        model = SimpleNamespace(
+            generation_config=SimpleNamespace(
+                eos_token_id=[248046, 248044],
+                pad_token_id=248044,
+            )
+        )
+        pipeline.handle = MagicMock(
+            architecture="qwen3_5", tokenizer=tokenizer, model=model
+        )
+
+        assert pipeline._deterministic_generation_kwargs(64) == {
+            "max_new_tokens": 64,
+            "do_sample": False,
+            "eos_token_id": [248046, 248044],
+            "pad_token_id": 248044,
+        }
+
+    def test_deterministic_generation_omits_absent_stop_tokens(self):
+        pipeline = AbliterationPipeline(model_name="bare-model")
+        tokenizer = SimpleNamespace(eos_token_id=None, pad_token_id=None)
+        pipeline.handle = SimpleNamespace(
+            tokenizer=tokenizer,
+            model=SimpleNamespace(generation_config=None),
+        )
+
+        assert pipeline._deterministic_generation_kwargs(12) == {
+            "max_new_tokens": 12,
+            "do_sample": False,
+        }
+
     def test_no_wrap_when_disabled(self):
         """Should not wrap prompts when use_chat_template is False."""
         pipeline = AbliterationPipeline(