Update README.md

README.md

@@ -10,8 +10,58 @@ license: mit
 |----|---------|---------|------|----|-------|--------|------|--------|
 |290k|   512   |  4096   |  8   | 32 |   4   |    2   |  96  | 229920 |
 
+"norm_eps": 1e-5
+
 仅做了自回归预训练，上下文长度512，没有监督微调等后训练过程，只用于验证下一词元生成的可行性。
 
 使用2025年12月构造的[4096词表](https://github.com/bd4sur/Nano/blob/master/tokenizer/tokenizer_4096.json)。
 
 Q80量化会完全损坏下一词元预测能力。
+
+## 训练参数
+
+230K
+
+```
+{
+    "use_lora": false,
+    "lora_rank": 8,
+    "lora_alpha": 16,
+    "lora_dropout": 0.0,
+
+    "from_checkpoint": "",
+    "save_checkpoint_to": "/home/bd4sur/ai/Nano/checkpoint",
+    "dataset_path": [
+        ["/home/bd4sur/ai/Nano/dataset_preprocessed/pt_train_0.base64", "/home/bd4sur/ai/Nano/dataset_preprocessed/pt_val_0.base64"]
+    ],
+    "tokenizer_path": "/home/bd4sur/ai/Nano/tokenizer/tokenizer_4096.json",
+
+    "random_seed": 39,
+    "batch_size": 256,
+    "gradient_accumulation_steps": 1,
+    "grad_clip": 1.0,
+
+    "dropout": 0.0,
+
+    "learning_rate": 5e-4,
+    "weight_decay": 1e-1,
+    "beta1": 0.9,
+    "beta2": 0.95,
+
+    "decay_lr": true,
+    "warmup_iters": 500,
+    "lr_decay_iters": 1e9,
+    "min_lr": 6e-5,
+
+    "eval_interval": 100,
+    "log_interval": 10,
+    "eval_iters": 5,
+
+    "backend": "nccl",
+    "device": "cuda",
+    "sdp_kernel": "flash",
+    "dtype": "bfloat16",
+    "use_amp": true
+}
+
+```