Upload vlm-streaming-sft-unsloth-qwen.py with huggingface_hub

vlm-streaming-sft-unsloth-qwen.py

@@ -296,21 +296,16 @@ def main():
     )
 
     # Convert streaming dataset to list (required for Qwen3-VL per Unsloth docs)
-    # "Using map kicks in dataset standardization which can be complicated"
     print("  Converting streaming dataset to list...")
     train_data = list(dataset.take(500))  # Take enough samples for training
     print(f"  Loaded {len(train_data)} samples")
 
-    # Get the actual tokenizer object (Qwen may return processor-like object)
-    actual_tokenizer = getattr(tokenizer, 'tokenizer', tokenizer)
-    print(f"  Tokenizer type: {type(tokenizer)}")
-    print(f"  Actual tokenizer type: {type(actual_tokenizer)}")
-
+    # Use older 'tokenizer=' parameter (not processing_class) - required for Unsloth VLM
     trainer = SFTTrainer(
         model=model,
-        train_dataset=train_data,
-        processing_class=actual_tokenizer,
+        tokenizer=tokenizer,  # Full processor, not processor.tokenizer
         data_collator=UnslothVisionDataCollator(model, tokenizer),
+        train_dataset=train_data,
         args=training_config,
     )