fix(base_trainer): force pad_token regardless of architecture

CarperAI · Jun 23, 2023 · 087e73d · 087e73d
1 parent f59086b
commit 087e73d
Showing 1 changed file with 2 additions and 3 deletions.
diff --git a/trlx/trainer/accelerate_base_trainer.py b/trlx/trainer/accelerate_base_trainer.py
@@ -73,9 +73,8 @@ def __init__(self, config, **kwargs):  # noqa: C901
         self.tokenizer.padding_side = config.tokenizer.padding_side
         self.tokenizer.truncation_side = config.tokenizer.truncation_side
         self.tokenizer.sep_token = "<sep>"
-        if config.model.model_arch_type != "seq2seq":
-            self.tokenizer.pad_token = self.tokenizer.eos_token
-            self.tokenizer.pad_token_id = self.tokenizer.eos_token_id
+        if self.tokenizer.pad_token is None:
+            self.tokenizer.pad_token = "<|padding|>"
 
         script_name = os.path.basename(sys.argv[0]).rsplit(".", 1)[0]
         if not isinstance(config.model.model_path, str):