SFT: 4 epochs larger curriculum

train_sft.py

@@ -37,14 +37,14 @@ def main() -> None:
         ),
         args=SFTConfig(
             output_dir=OUTPUT_DIR,
-            num_train_epochs=3,
+            num_train_epochs=4,
             per_device_train_batch_size=2,
             per_device_eval_batch_size=2,
             gradient_accumulation_steps=8,
             learning_rate=2e-4,
-            logging_steps=5,
+            logging_steps=10,
             eval_strategy="steps",
-            eval_steps=20,
+            eval_steps=40,
             save_strategy="epoch",
             save_total_limit=1,
             max_length=1024,