🧠 两小时从 0 训出你的 64M 大模型。Train your own 64M LLM from scratch in 2 hours.
update lr
gongjybab480073ed6
diff --git a/train_full_sft.py b/train_full_sft.pyindex 859dafc..5185861 100644--- a/train_full_sft.py+++ b/train_full_sft.py@@ -122,7 +122,7 @@ def init_distributed_mode():if __name__ == "__main__":parser = argparse.ArgumentParser(description="MiniMind Full SFT")parser.add_argument("--out_dir", type=str, default="out")- parser.add_argument("--epochs", type=int, default=6)+ parser.add_argument("--epochs", type=int, default=1)parser.add_argument("--batch_size", type=int, default=32)parser.add_argument("--learning_rate", type=float, default=5e-5)parser.add_argument("--device", type=str, default="cuda:0" if torch.cuda.is_available() else "cpu")