change to right-padding, update reward score #803

2025-12-15 19:30:36 +08:00 · 2023-09-08 20:04:31 +08:00
parent 8aaaa132d4
commit 8ea32e4046
15 changed files with 97 additions and 59 deletions
--- a/src/llmtuner/tuner/core/loader.py
+++ b/src/llmtuner/tuner/core/loader.py
@@ -68,7 +68,7 @@ def load_model_and_tokenizer(
    tokenizer = AutoTokenizer.from_pretrained(
        model_args.model_name_or_path,
        use_fast=model_args.use_fast_tokenizer,
-        padding_side=model_args.padding_side,
+        padding_side="right", # training with left-padded tensors in fp16 precision may cause overflow
        **config_kwargs
    )