Riyuechang
commited on
Commit
•
333b089
1
Parent(s):
16e574d
Update README.md
Browse files
README.md
CHANGED
@@ -1,3 +1,58 @@
|
|
1 |
-
---
|
2 |
-
license: gemma
|
3 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
---
|
2 |
+
license: gemma
|
3 |
+
base_model:
|
4 |
+
- google/gemma-2-2b-it
|
5 |
+
datasets:
|
6 |
+
- Riyuechang/PTT-Corpus-100K_Gossiping-1400-39400_v2
|
7 |
+
pipeline_tag: text-generation
|
8 |
+
tags:
|
9 |
+
- PTT
|
10 |
+
- PTT_Chat
|
11 |
+
---
|
12 |
+
|
13 |
+
# 版本資訊
|
14 |
+
使用新的噪聲較小(理論上)的數據訓練
|
15 |
+
Lora使用了更大的r(32)
|
16 |
+
取消了Dora
|
17 |
+
因為Dora的提升有限,還會大幅降低訓練和推理的效率
|
18 |
+
|
19 |
+
# 簡介
|
20 |
+
本模型是基於[google/gemma-2-2b-it](https://huggingface.co/google/gemma-2-2b-it)微調後的產物
|
21 |
+
模型使用來自[PTT](https://www.ptt.cc/bbs/index.html)網站中的[Gossiping](https://www.ptt.cc/bbs/Gossiping/index.html)分類的資料訓練
|
22 |
+
過程中使用了一些方法從海量的數據中,過濾出噪聲較小(理論上)的部份作為訓練數據
|
23 |
+
訓練資料: [Riyuechang/PTT-Corpus-100K_Gossiping-1400-39400_v2](https://huggingface.co/datasets/Riyuechang/PTT-Corpus-100K_Gossiping-1400-39400_v2)
|
24 |
+
|
25 |
+
# 設備
|
26 |
+
- Ubuntu 22.04.4 LTS
|
27 |
+
- NVIDIA GeForce RTX 3060 12G
|
28 |
+
|
29 |
+
# Lora參數
|
30 |
+
```python
|
31 |
+
r=32,
|
32 |
+
lora_alpha=32,
|
33 |
+
lora_dropout=0.1,
|
34 |
+
task_type="CAUSAL_LM",
|
35 |
+
target_modules="all-linear",
|
36 |
+
bias="none",
|
37 |
+
use_rslora=True
|
38 |
+
```
|
39 |
+
|
40 |
+
# 訓練參數
|
41 |
+
```python
|
42 |
+
per_device_train_batch_size=28,
|
43 |
+
gradient_accumulation_steps=1,
|
44 |
+
num_train_epochs=3,
|
45 |
+
warmup_ratio=0.1,
|
46 |
+
learning_rate=5e-5,
|
47 |
+
bf16=True,
|
48 |
+
save_strategy="steps",
|
49 |
+
save_steps=1000,
|
50 |
+
save_total_limit=5,
|
51 |
+
logging_steps=10,
|
52 |
+
output_dir=log_output,
|
53 |
+
optim="paged_adamw_8bit",
|
54 |
+
gradient_checkpointing=True
|
55 |
+
```
|
56 |
+
|
57 |
+
# 結果
|
58 |
+
- loss: 1.8808
|