Looks like the decoder start token should actually be 0

README.md

@@ -447,7 +447,7 @@ input_ids = tokenizer(text, return_tensors="pt").input_ids
 outputs = model.generate(
     input_ids=input_ids,
     generation_config=GenerationConfig(
-        decoder_start_token_id=1,
+        decoder_start_token_id=0,
 ))
 
 tokenizer.decode(outputs[0], skip_special_tokens=True)