Add check that the prompt is the same for the PREFERENCE dataset format

Author: VProv

src/together/utils/files.py

@@ -304,6 +304,31 @@ def _check_jsonl(file: Path) -> Dict[str, Any]:
                                 line_number=idx + 1,
                                 error_source="key_value",
                             )
+                    # Check that all messages except the last one are the same for "chosen" and "rejected"
+                    chosen_messages = json_line["chosen"]
+                    rejected_messages = json_line["rejected"]
+
+                    if len(chosen_messages) != len(rejected_messages):
+                        raise InvalidFileFormatError(
+                            message="The 'chosen' and 'rejected' lists must have the same number of messages.",
+                            line_number=idx + 1,
+                            error_source="key_value",
+                        )
+
+                    # Count discrepancies between messages using a generator
+                    discrepancies = sum(
+                        1
+                        for i in range(len(chosen_messages) - 1)
+                        if chosen_messages[i] != rejected_messages[i]
+                    )
+
+                    if discrepancies > 1:
+                        raise InvalidFileFormatError(
+                            message=f"Found {discrepancies} different messages between 'chosen' and 'rejected'. "
+                            "Only the last message should differ.",
+                            line_number=idx + 1,
+                            error_source="key_value",
+                        )
                 elif current_format == DatasetFormat.CONVERSATION:
                     message_column = JSONL_REQUIRED_COLUMNS_MAP[
                         DatasetFormat.CONVERSATION