Update README.md
Browse files
README.md
CHANGED
@@ -28,12 +28,15 @@ pipeline_tag: text-generation
|
|
28 |
|:-------:|:-------:|:-------:|:-------:|
|
29 |
| 20230827-v0.1 | 🤗 [Sakura-13B-Galgame-v0.1](https://huggingface.co/sakuraumi/Sakura-13B-Galgame/tree/main/sakura_13b_model_v0.1) | - | - |
|
30 |
| 20230908-v0.4 | 🤗 [Sakura-13B-Galgame-v0.4](https://huggingface.co/sakuraumi/Sakura-13B-Galgame/tree/main/sakura_13b_model_v0.4) | - | - |
|
31 |
-
| 20230917-v0.5 | 🤗 sakuraumi/Sakura-13B-Galgame根目录下 | 🤗 [Sakura-13B-Galgame-v0.5-8bits](https://huggingface.co/sakuraumi/Sakura-13B-Galgame/tree/main/sakura_13b_model_v0.5_8bits) | - |
|
|
|
32 |
|
33 |
目前仍为实验版本,翻译质量较差.
|
34 |
|
35 |
# 日志
|
36 |
|
|
|
|
|
37 |
`20230917`:上传第三版模型`sakura-13b-2epoch-2.6M-0917-v0.5`,改用Baichuan2-13B-Chat模型进行微调,翻译质量有所提高。
|
38 |
|
39 |
`20230908`:上传第二版模型`sakura-13b-1epoch-2.6M-0903-v0.4`,使用Galgame和轻小说数据集进行微调,语法能力有所提高。轻小说数据集感谢[CjangCjengh](https://github.com/CjangCjengh)大佬提供。
|
@@ -51,16 +54,33 @@ pipeline_tag: text-generation
|
|
51 |
- Base model: [LLaMA2-13B](https://huggingface.co/meta-llama/Llama-2-13b-chat-hf)
|
52 |
- Languages: Chinese/Japanese
|
53 |
|
54 |
-
### v0.5
|
55 |
|
56 |
- Finetuned by [SakuraUmi](https://github.com/pipixia244)
|
57 |
- Finetuned on [Baichuan2-13B-Chat](https://huggingface.co/baichuan-inc/Baichuan2-13B-Chat)
|
58 |
- Base model: [Baichuan2-13B-Base](https://huggingface.co/baichuan-inc/Baichuan2-13B-Base)
|
59 |
- Languages: Chinese/Japanese
|
60 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
61 |
## 版本
|
62 |
|
63 |
-
### v0.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
64 |
|
65 |
数据集:约260k Galgame中日文本 + 约2.3M 轻小说中日文本
|
66 |
|
@@ -131,26 +151,29 @@ pipeline_tag: text-generation
|
|
131 |
prompt = "<reserved_106>" + query + "<reserved_107>"
|
132 |
```
|
133 |
|
134 |
-
-
|
|
|
|
|
135 |
|
136 |
-
|
137 |
-
| ---- | ---- | --- |
|
138 |
-
| beams number | 2 | 1 |
|
139 |
-
| repetition penalty | 1 | 1 |
|
140 |
-
| max new token | 512 | 512 |
|
141 |
-
| min new token | 1 | - |
|
142 |
|
143 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
144 |
|
145 |
- 量化:
|
146 |
|
147 |
-
|
148 |
-
|
149 |
-
其余推理流程与LLaMA2(v0.1-v0.4)/Baichuan2(v0.5+)一致
|
150 |
|
151 |
# 微调
|
152 |
|
153 |
-
流程与LLaMA2(v0.1-v0.4)/Baichuan2(v0.5+)一致,prompt构造参考推理部分
|
154 |
|
155 |
# 后续工作
|
156 |
|
|
|
28 |
|:-------:|:-------:|:-------:|:-------:|
|
29 |
| 20230827-v0.1 | 🤗 [Sakura-13B-Galgame-v0.1](https://huggingface.co/sakuraumi/Sakura-13B-Galgame/tree/main/sakura_13b_model_v0.1) | - | - |
|
30 |
| 20230908-v0.4 | 🤗 [Sakura-13B-Galgame-v0.4](https://huggingface.co/sakuraumi/Sakura-13B-Galgame/tree/main/sakura_13b_model_v0.4) | - | - |
|
31 |
+
| 20230917-v0.5 | 🤗 sakuraumi/Sakura-13B-Galgame根目录下 | 🤗 [Sakura-13B-Galgame-v0.5-8bits](https://huggingface.co/sakuraumi/Sakura-13B-Galgame/tree/main/sakura_13b_model_v0.5_8bits) | [Sakura-13B-Galgame-v0.5-4bits](https://huggingface.co/sakuraumi/Sakura-13B-Galgame/tree/main/sakura_13b_model_v0.5_4bits_autogptq_40k) |
|
32 |
+
| 20231011-v0.7 | 🤗 [Sakura-14B-LNovel](https://huggingface.co/sakuraumi/Sakura-14B-LNovel)
|
33 |
|
34 |
目前仍为实验版本,翻译质量较差.
|
35 |
|
36 |
# 日志
|
37 |
|
38 |
+
`20231011`:上传第四版模型`sakura-14b-2epoch-4.4M-1003-v0.7`,改用QWen-14B-Chat模型进行微调,增加数据集。
|
39 |
+
|
40 |
`20230917`:上传第三版模型`sakura-13b-2epoch-2.6M-0917-v0.5`,改用Baichuan2-13B-Chat模型进行微调,翻译质量有所提高。
|
41 |
|
42 |
`20230908`:上传第二版模型`sakura-13b-1epoch-2.6M-0903-v0.4`,使用Galgame和轻小说数据集进行微调,语法能力有所提高。轻小说数据集感谢[CjangCjengh](https://github.com/CjangCjengh)大佬提供。
|
|
|
54 |
- Base model: [LLaMA2-13B](https://huggingface.co/meta-llama/Llama-2-13b-chat-hf)
|
55 |
- Languages: Chinese/Japanese
|
56 |
|
57 |
+
### v0.5
|
58 |
|
59 |
- Finetuned by [SakuraUmi](https://github.com/pipixia244)
|
60 |
- Finetuned on [Baichuan2-13B-Chat](https://huggingface.co/baichuan-inc/Baichuan2-13B-Chat)
|
61 |
- Base model: [Baichuan2-13B-Base](https://huggingface.co/baichuan-inc/Baichuan2-13B-Base)
|
62 |
- Languages: Chinese/Japanese
|
63 |
|
64 |
+
### v0.7
|
65 |
+
|
66 |
+
- Finetuned by [SakuraUmi](https://github.com/pipixia244)
|
67 |
+
- Finetuned on [Qwen-14B-Chat](https://huggingface.co/Qwen/Qwen-14B)
|
68 |
+
- Base model: [Qwen-14B](https://huggingface.co/Qwen/Qwen-14B)
|
69 |
+
- Languages: Chinese/Japanese
|
70 |
+
|
71 |
## 版本
|
72 |
|
73 |
+
### v0.7
|
74 |
+
|
75 |
+
数据集:约1M Galgame中日文本 + 约3.4M 轻小说中日文本
|
76 |
+
|
77 |
+
微调方式:全参数
|
78 |
+
|
79 |
+
微调epoch数:2
|
80 |
+
|
81 |
+
参数量:14B
|
82 |
+
|
83 |
+
### v0.5
|
84 |
|
85 |
数据集:约260k Galgame中日文本 + 约2.3M 轻小说中日文本
|
86 |
|
|
|
151 |
prompt = "<reserved_106>" + query + "<reserved_107>"
|
152 |
```
|
153 |
|
154 |
+
- v0.7
|
155 |
+
参考Qwen-14B-Chat的prompt构造方式:[这里](https://huggingface.co/Qwen/Qwen-14B-Chat/blob/5188dfeb4ff175705aa3a84ef9d616c70dea029b/qwen_generation_utils.py#L119)和[这里](https://github.com/hiyouga/LLaMA-Efficient-Tuning/blob/5310e4d1829f36619c8f224d09ec15eeaf7a4877/src/llmtuner/extras/template.py#L546)
|
156 |
+
|
157 |
|
158 |
+
- 推理与解码参数:
|
|
|
|
|
|
|
|
|
|
|
159 |
|
160 |
+
| 参数 | 值 |
|
161 |
+
| ---- | ---- |
|
162 |
+
| temperature | 1 |
|
163 |
+
| top p | 0.5 |
|
164 |
+
| do sample | True |
|
165 |
+
| beams number | 1 |
|
166 |
+
| repetition penalty | 1 |
|
167 |
+
| max new token | 512 |
|
168 |
+
| min new token | 1 |
|
169 |
|
170 |
- 量化:
|
171 |
|
172 |
+
根据transformers文档中给出的AutoGPTQ量化教程自行量化,或使用我们已经量化好的模型。
|
|
|
|
|
173 |
|
174 |
# 微调
|
175 |
|
176 |
+
流程与LLaMA2(v0.1-v0.4)/Baichuan2(v0.5+)/Qwen14B(v0.7)一致,prompt构造参考推理部分
|
177 |
|
178 |
# 后续工作
|
179 |
|