.

train_sft_ds.py

@@ -3,6 +3,7 @@ import os
 import glob
 import socket
 import argparse
+import inspect
 from typing import Dict, List, Iterable, Iterator, Tuple, Optional
 
 import torch
@@ -217,7 +218,7 @@ def parse_args():
     ap.add_argument("--eval_data_glob", type=str, default=None,
                     help="(可选) 测试/验证集 jsonl 通配符；如提供则优先使用")
     ap.add_argument("--local_rank", type=int, default=-1,
-                    help="for deepspeed/torchrun launcher; ignored by user code")    
+                    help="for deepspeed/torchrun launcher; ignored by user code")
     return ap.parse_args()
 
 
@@ -341,12 +342,26 @@ def main():
     logging_dir = os.path.join(args.output_dir, "logs")
     os.makedirs(logging_dir, exist_ok=True)
 
+    # ---- 兼容 4.51（eval_strategy）与旧版（evaluation_strategy） ----
+    ta_kwargs = {}
+    sig = inspect.signature(TrainingArguments.__init__).parameters
+    if eval_dataset is not None:
+        if "eval_strategy" in sig:
+            ta_kwargs["eval_strategy"] = "steps"
+        elif "evaluation_strategy" in sig:
+            ta_kwargs["evaluation_strategy"] = "steps"
+    else:
+        if "eval_strategy" in sig:
+            ta_kwargs["eval_strategy"] = "no"
+        elif "evaluation_strategy" in sig:
+            ta_kwargs["evaluation_strategy"] = "no"
+
     training_args = TrainingArguments(
         output_dir=args.output_dir,
         logging_dir=logging_dir,
         do_train=True,
         do_eval=(eval_dataset is not None),
-        evaluation_strategy=("steps" if eval_dataset is not None else "no"),
+        # evaluation_strategy / eval_strategy 通过 **ta_kwargs 传入
         eval_steps=max(50, args.save_steps // 5) if eval_dataset is not None else None,
         per_device_train_batch_size=args.per_device_train_batch_size,
         gradient_accumulation_steps=args.gradient_accumulation_steps,
@@ -368,7 +383,8 @@ def main():
         gradient_checkpointing=args.gradient_checkpointing,
         remove_unused_columns=False,   # 需要保留我们的字段
         torch_compile=False,
-        save_on_each_node=False
+        save_on_each_node=False,
+        **ta_kwargs,  # ← 兼容参数
     )
 
     trainer = Trainer(