fix: use args.context_length for packing, lower save_steps

Author: walidsobhie-code

training/train_extended_context.py

@@ -291,8 +291,8 @@ def tokenize(example):
         gradient_checkpointing=True,
         gradient_checkpointing_kwargs={"use_reentrant": False},
         logging_steps=10,
-        save_steps=100,
-        eval_steps=100,
+        save_steps=50,
+        eval_steps=50,
         save_total_limit=2,
         optim="adamw_torch",
         report_to="none",