Update README.md
Browse files
README.md
CHANGED
|
@@ -2,25 +2,20 @@
|
|
| 2 |
license: mit
|
| 3 |
---
|
| 4 |
|
| 5 |
-
# Nano语言模型 -
|
| 6 |
|
| 7 |
用于测试的极小语言模型，有以下几个规格。
|
| 8 |
|
| 9 |
-
| |BlockSize|VocabSize|Layers|Embd|Q_Heads|KV_Heads|Hidden| #Param |
|
| 10 |
-
|----|---------|---------|------|----|-------|--------|------|--------|
|
| 11 |
-
|
|
| 12 |
-
|
| 13 |
-
"norm_eps": 1e-5
|
| 14 |
-
|
| 15 |
-
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。
|
| 16 |
|
| 17 |
使用2025年12月构造的[4096词表](https://github.com/bd4sur/Nano/blob/master/tokenizer/tokenizer_4096.json)。
|
| 18 |
|
| 19 |
-
Q80量化会完全损坏下一词元预测能力。
|
| 20 |
-
|
| 21 |
## 训练参数
|
| 22 |
|
| 23 |
-
|
| 24 |
|
| 25 |
```
|
| 26 |
{
|
|
@@ -63,5 +58,49 @@ Q80量化会完全损坏下一词元预测能力。
|
|
| 63 |
"dtype": "bfloat16",
|
| 64 |
"use_amp": true
|
| 65 |
}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 66 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 67 |
```
|
|
|
|
| 2 |
license: mit
|
| 3 |
---
|
| 4 |
|
| 5 |
+
# Nano语言模型 - 极小模型系列
|
| 6 |
|
| 7 |
用于测试的极小语言模型，有以下几个规格。
|
| 8 |
|
| 9 |
+
| Model Name |BlockSize|VocabSize|Layers| Embd |Q_Heads|KV_Heads|Hidden|NormEps| #Param |
|
| 10 |
+
|-----------------|---------|---------|------|------|-------|--------|------|-------|--------|
|
| 11 |
+
|Psycho-230k-base | 512 | 4096 | 8 | 32 | 4 | 2 | 96 | 1e-5 | 229920 |
|
| 12 |
+
|Nano-230k-base | 512 | 4096 | 8 | 32 | 4 | 2 | 96 | 1e-5 | 229920 |
|
|
|
|
|
|
|
|
|
|
| 13 |
|
| 14 |
使用2025年12月构造的[4096词表](https://github.com/bd4sur/Nano/blob/master/tokenizer/tokenizer_4096.json)。
|
| 15 |
|
|
|
|
|
|
|
| 16 |
## 训练参数
|
| 17 |
|
| 18 |
+
Psycho-230k-base
|
| 19 |
|
| 20 |
```
|
| 21 |
{
|
|
|
|
| 58 |
"dtype": "bfloat16",
|
| 59 |
"use_amp": true
|
| 60 |
}
|
| 61 |
+
```
|
| 62 |
+
|
| 63 |
+
Nano-230k-base
|
| 64 |
+
|
| 65 |
+
```
|
| 66 |
+
{
|
| 67 |
+
"use_lora": false,
|
| 68 |
+
"lora_rank": 8,
|
| 69 |
+
"lora_alpha": 16,
|
| 70 |
+
"lora_dropout": 0.0,
|
| 71 |
+
|
| 72 |
+
"from_checkpoint": "",
|
| 73 |
+
"save_checkpoint_to": "/home/bd4sur/ai/Nano/checkpoint",
|
| 74 |
+
"dataset_path": [
|
| 75 |
+
["/home/bd4sur/ai/Nano/dataset_preprocessed/pt_1Gtk_512_4096_train.base64", "/home/bd4sur/ai/Nano/dataset_preprocessed/pt_1Gtk_512_4096_valid.base64"]
|
| 76 |
+
],
|
| 77 |
+
"tokenizer_path": "/home/bd4sur/ai/Nano/tokenizer/tokenizer_4096.json",
|
| 78 |
+
|
| 79 |
+
"random_seed": 39,
|
| 80 |
+
"batch_size": 256,
|
| 81 |
+
"gradient_accumulation_steps": 1,
|
| 82 |
+
"grad_clip": 1.0,
|
| 83 |
|
| 84 |
+
"dropout": 0.0,
|
| 85 |
+
|
| 86 |
+
"learning_rate": 5e-4,
|
| 87 |
+
"weight_decay": 1e-1,
|
| 88 |
+
"beta1": 0.9,
|
| 89 |
+
"beta2": 0.95,
|
| 90 |
+
|
| 91 |
+
"decay_lr": true,
|
| 92 |
+
"warmup_iters": 500,
|
| 93 |
+
"lr_decay_iters": 1e9,
|
| 94 |
+
"min_lr": 6e-5,
|
| 95 |
+
|
| 96 |
+
"eval_interval": 100,
|
| 97 |
+
"log_interval": 10,
|
| 98 |
+
"eval_iters": 5,
|
| 99 |
+
|
| 100 |
+
"backend": "nccl",
|
| 101 |
+
"device": "cuda",
|
| 102 |
+
"sdp_kernel": "flash",
|
| 103 |
+
"dtype": "bfloat16",
|
| 104 |
+
"use_amp": true
|
| 105 |
+
}
|
| 106 |
```
|