Merge pull request #5115 from YeQiuO/main

fix: `Train on the last turn only` truncate bug

src/llamafactory/data/loader.py

@@ -206,6 +206,8 @@ def get_dataset(
     template = get_template_and_fix_tokenizer(tokenizer, data_args.template, data_args.tool_format)
     if data_args.train_on_prompt and template.efficient_eos:
         raise ValueError("Current template does not support `train_on_prompt`.")
+    if stage!="sft" and data_args.mask_history:
+        raise ValueError("`Train on the last turn only` is only valid for sft training.")
 
     # Load tokenized dataset
     if data_args.tokenized_path is not None:

src/llamafactory/data/processors/supervised.py

@@ -53,7 +53,7 @@ def _encode_supervised_example(
         input_ids += [image_token_id] * getattr(processor, "image_seq_length")
         labels += [IGNORE_INDEX] * getattr(processor, "image_seq_length")
 
-    encoded_pairs = template.encode_multiturn(tokenizer, messages, system, tools)
+    encoded_pairs = template.encode_multiturn(tokenizer, messages, system, tools, mask_history)
     total_length = 1 if template.efficient_eos else 0
     for turn_idx, (source_ids, target_ids) in enumerate(encoded_pairs):
         if total_length >= cutoff_len:
@@ -71,14 +71,15 @@ def _encode_supervised_example(
         else:
             source_label = [IGNORE_INDEX] * source_len
 
-        if mask_history and turn_idx != len(encoded_pairs) - 1:
-            target_label = [IGNORE_INDEX] * target_len
+        if mask_history:
+            target_label = target_ids if turn_idx==0 else [IGNORE_INDEX] * target_len
+            input_ids = source_ids + target_ids + input_ids
+            labels = source_label + target_label + labels
         else:
             target_label = target_ids
-
-        input_ids += source_ids + target_ids
-        labels += source_label + target_label
-
+            input_ids += source_ids + target_ids
+            labels += source_label + target_label
+            
     if template.efficient_eos:
         input_ids += [tokenizer.eos_token_id]
         labels += [tokenizer.eos_token_id]

src/llamafactory/data/template.py

@@ -69,12 +69,16 @@ class Template:
         messages: Sequence[Dict[str, str]],
         system: Optional[str] = None,
         tools: Optional[str] = None,
+        mask_history: bool = False,
     ) -> List[Tuple[List[int], List[int]]]:
         r"""
         Returns multiple pairs of token ids representing prompts and responses respectively.
         """
         encoded_messages = self._encode(tokenizer, messages, system, tools)
-        return [(encoded_messages[i], encoded_messages[i + 1]) for i in range(0, len(encoded_messages), 2)]
+        if not mask_history:
+            return [(encoded_messages[i], encoded_messages[i + 1]) for i in range(0, len(encoded_messages), 2)]
+        else:
+            return [(encoded_messages[i], encoded_messages[i + 1]) for i in range(len(encoded_messages)-2, -1, -2)]
 
     def extract_tool(self, content: str) -> Union[str, List[Tuple[str, str]]]:
         r"""

src/llamafactory/hparams/data_args.py

@@ -141,3 +141,6 @@ class DataArguments:
 
         if self.streaming and self.max_samples is not None:
             raise ValueError("`max_samples` is incompatible with `streaming`.")
+
+        if self.mask_history and self.train_on_prompt:
+            raise ValueError("`Train on the last turn only` does not support `train_on_prompt`.")