fix: match eval and save strategy (both steps/4000)

train_hrm_text_pi.py

@@ -928,7 +928,8 @@ def main():
         weight_decay=0.01,
         warmup_steps=500 if not args.test else 0,
         lr_scheduler_type="cosine",
-        eval_strategy="epoch",
+        eval_strategy="steps",
+        eval_steps=4000,
         save_strategy="steps",
         save_steps=4000,
         load_best_model_at_end=True,