fix: Qwen3.5 VL processor has no .encode() — use .tokenizer attr

train_v4.py

@@ -97,6 +97,8 @@ def fix_tool_calls(messages):
 def load_and_format(path):
     """Load JSONL manually — pyarrow chokes on mixed tool_calls argument types."""
     from datasets import Dataset
+    # Qwen3.5 is a VL model — from_pretrained returns a Processor, not a Tokenizer
+    _enc = tokenizer.tokenizer if hasattr(tokenizer, 'tokenizer') else tokenizer
     texts = []
     skipped = 0
     with open(path) as f:
@@ -119,7 +121,7 @@ def load_and_format(path):
                     tokenize=False,
                     add_generation_prompt=False,
                 )
-            if len(tokenizer.encode(text)) <= MAX_SEQ:
+            if len(_enc.encode(text)) <= MAX_SEQ:
                 texts.append(text)
             else:
                 skipped += 1