fix: 修复强化学习算法问题

2026-03-19 22:23:51 +08:00
parent a5574f92e2
commit 0f518473af
2 changed files with 62 additions and 83 deletions
@@ -112,11 +112,8 @@ def train(

    model = parameter.model
    
-    kwargs = {
+    strategy_kwargs = {
        "dpo_beta": dpo_beta,
-        "bos_token_id": parameter.tokenizer.bos_id,
-        "eos_token_id": parameter.tokenizer.eos_id,
-        "pad_token_id": parameter.tokenizer.pad_id,
        "label_smoothing": label_smoothing
    }

@@ -158,7 +155,7 @@ def train(
        parallel_wrapper=ddp_wrap,
        state_dict_fn=prepare_checkpoint,
        device_type=device_type,
-        extra_kwargs=kwargs,
+        extra_kwargs=strategy_kwargs,
    )
    
    trainer = Trainer(train_config)