Update train_model.py

train_model.py

@@ -111,25 +111,25 @@ data_collator = DataCollatorForSeq2Seq(tokenizer, model=model, padding=True)
 training_args = TrainingArguments(
     output_dir="./results",
     eval_strategy="steps",
-    eval_steps=250,  # 🔹 Avaliação frequente
+    eval_steps=1000,  # 🔹 Avaliação frequente
     save_strategy="steps",
-    save_steps=500,  # 🔹 Salvar checkpoints
+    save_steps=2500,  # 🔹 Salvar checkpoints
     per_device_train_batch_size=2,
     per_device_eval_batch_size=2,
-    gradient_accumulation_steps=6,
+    gradient_accumulation_steps=4,
     num_train_epochs=10,  # 🔹 Mais épocas para um treinamento robusto
     weight_decay=0.01,
     logging_dir="./logs",
     logging_strategy="steps",
-    logging_steps=75,  # 🔹 Logs frequentes
-    save_total_limit=5,
+    logging_steps=50,  # 🔹 Logs frequentes
+    save_total_limit=2,
     push_to_hub=True,
     hub_model_id=NEW_MODEL_NAME,
     hub_token=HF_API_KEY,
     gradient_checkpointing=True,
     bf16=True,
     learning_rate=1e-6,  # 🔹 Aprendizado mais lento
-    max_grad_norm=0.5,
+    max_grad_norm=0.2,
     warmup_ratio=0.1,
     lr_scheduler_type="cosine",
     optim="adamw_torch"