[data] fix template (#8827)

src/llamafactory/data/processor/supervised.py

@@ -62,7 +62,7 @@ class SupervisedDatasetProcessor(DatasetProcessor):
 
             if self.data_args.train_on_prompt:
                 source_label = source_ids
-            elif self.template.efficient_eos:
+            elif self.template.efficient_eos and turn_idx != 0:
                 source_label = [self.tokenizer.eos_token_id] + [IGNORE_INDEX] * (source_len - 1)
             else:
                 source_label = [IGNORE_INDEX] * source_len

src/llamafactory/data/template.py

@@ -1069,7 +1069,9 @@ register_template(
     format_assistant=StringFormatter(slots=["{{content}}<|end|>"]),
     format_system=StringFormatter(slots=["<|start|>system<|message|>{{content}}<|end|>"]),
     default_system="You are ChatGPT, a large language model trained by OpenAI.",
+    thought_words=("<|channel|>analysis<|message|>", "<|end|><|start|>assistant<|channel|>final<|message|>"),
     efficient_eos=True,
+    template_class=ReasoningTemplate,
 )