File size: 9,620 Bytes
9255e5c
 
 
 
 
 
 
442a193
 
 
9255e5c
 
442a193
9255e5c
 
bed2d73
 
9255e5c
c104028
 
 
 
9255e5c
 
442a193
 
 
 
f96f4cb
442a193
9255e5c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
efd737b
9255e5c
 
 
 
f674f65
9255e5c
 
 
f674f65
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5359d7e
9255e5c
 
 
f674f65
9255e5c
 
 
 
 
 
f674f65
9255e5c
 
d053eb5
f68c20c
bbf5526
 
f674f65
 
bbf5526
 
 
 
 
 
f68c20c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
d053eb5
 
 
 
 
 
 
 
 
 
f674f65
bbf5526
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
d053eb5
 
 
 
 
 
 
f674f65
8c70874
 
f674f65
815696a
f674f65
9255e5c
 
 
 
 
 
f57ce0a
9255e5c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
bc66a0e
 
 
9255e5c
 
96709dd
9255e5c
 
 
e5daf15
 
96709dd
9255e5c
50ff845
e5daf15
 
 
 
 
 
 
9255e5c
 
 
 
652f4a4
9255e5c
652f4a4
9255e5c
 
 
 
bed2d73
 
 
930b7a0
bed2d73
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
---
license: apache-2.0
language:
- vi
- en
---

<p align="center">
  <img src="https://cdn-uploads.huggingface.co/production/uploads/63905e87df447b438817b2cd/QFhLKQlWeyO9XumtyghVo.jpeg" alt="Image" style="width: 400px; height: auto; border-radius: 10px;" />
</p>


## Model Details

- **Developed by:** Tuan Pham (FPTU HCM Student)
  - Contact me at: weekend.2810@gmail.com or tuanpmse160561@fpt.edu.vn
  - Looking for intern opportunity :D 
- **Model type:** Llama2-7B Decoder-only
- **Finetuned from model :**
  * meta-llama/Llama-2-7b
  * bkai-foundation-models/vietnamese-llama2-7b-120GB
  * yeen214/llama2_7b_merge_orcafamily.
- **Bilingual support :** English and Vietnamese

### Model Description

<!-- Provide a longer summary of what this model is. -->

This model is a proof of effort that one man can fine-tune his own model to reach SOTA.

### Model Sources

<!-- Provide the basic links for the model. -->

- **Repository:** 
  * Training: https://github.com/vTuanpham/Vietnamese_QA_System
  * Data: https://github.com/vTuanpham/Large_dataset_translator
- **Paper:** ...
- **Demo:** ...

## Uses

<!-- Address questions around how the model is intended to be used, including the foreseeable users of the model and those affected by the model. -->

### Prompt template

```
[SYSTEM_PROMPT]

 ####### Instruction:
[INPUT]

 %%%%%%% Response:
[RESPONSE]
```
Recommend keeping the system prompt in english.
## How to Get Started with the Model

Use the code below to get started with the model.
```python
import torch
from torch.cuda.amp import autocast
from transformers import AutoModelForCausalLM, AutoTokenizer, TextStreamer, pipeline


def prompt_format(system_prompt, instruction):
    prompt = f"""{system_prompt}

 ####### Instruction:
{instruction}

 %%%%%%% Response:
"""
    return prompt

system_prompt = """
You're an AI Large Language Model developed(created) by an AI developer named Tuấn, the architecture of you is decoder-based LM, your task are to think loudly step by step before give a good and relevant response
to the user request, answer in the language the user preferred.

The AI has been trained to answer questions, provide recommendations, and help with decision making. The AI thinks outside the box and follows the user requests
"""
instruction = "Xin chào"

formatted_prompt = prompt_format(system_prompt, instruction)
print(formatted_prompt)

model_name = "1TuanPham/T-Llama"
model = AutoModelForCausalLM.from_pretrained(model_name,
                                             torch_dtype=torch.bfloat16,
                                             use_cache=True,
                                             device_map="auto"
                                             )
tokenizer = AutoTokenizer.from_pretrained(model_name, use_fast=True)
streamer = TextStreamer(tokenizer, skip_special_tokens=True)
pipe = pipeline("text-generation", model=base_model, tokenizer=tokenizer, streamer=streamer)

with autocast():
  output_default = pipe(formatted_prompt, pad_token_id=50256, max_new_tokens=128)

```
Example outputs:
```
> Xin chào

Xin chào! Tôi là một AI được phát triển bởi một AI nhà phát triển tên là Tuấn. Tôi được thiết kế để giúp đỡ người dùng bằng cách trả lời các câu hỏi, đưa ra đề xuất và hỗ trợ trong quá trình ra quyết định.
Tôi có thể hỗ trợ bạn bằng cách nghĩ ra các câu trả lời hay và phù hợp cho các câu hỏi của bạn.

> Việt Nam có cà phê gì nổi tiếng thế giới ?

Cà phê robusta của Việt Nam nổi tiếng trên toàn cầu. Loại cà phê này được trồng chủ yếu ở các vùng đất cao như Lâm Đồng, Gia Lai, KonTum và có hương vị đậm đà, đắng hơn so với cà phê Arabica.
Cà phê robusta thường được sử dụng để pha chế các loại cà phê espresso và phin truyền thống.

> Tôi có 500$ hãy build cho tôi một bộ PC

Để xây dựng một bộ PC với 500 USD, bạn có thể bắt đầu bằng cách chọn các thành phần chính như CPU, GPU, bộ nhớ và bộ lưu trữ. Sau đó, bạn có thể thêm các thành phần bổ sung như loa, bàn phím và chuột để hoàn thiện hệ thống.

Dưới đây là một số gợi ý về cách bạn có thể cấu thành một bộ PC với 500 USD:

1. CPU: Intel Core i5-12400 hoặc AMD Ryzen 5 5600X là những lựa chọn tốt để bắt đầu.

2. GPU: Nvidia GeForce GTX 1650 hoặc AMD Radeon RX 570 là những lựa chọn tốt để bắt đầu.

3. Bộ nhớ: 8GB RAM là một khởi đầu tốt.

4. Bộ lưu trữ: 500 GB ổ cứng SSD là một khởi đầu tốt.

5. Bàn phím và chuột: Bàn phím và chuột có thể được mua riêng hoặc cùng một bộ.

6. Loa: Một loa có thể được mua riêng hoặc cùng một bộ với bàn phím và chuột.

> Cho tôi 3 thức uống để giúp ngủ ngon hơn

Chắc chắn! Dưới đây là ba thức uống có thể giúp bạn ngủ ngon hơn:

1. Mật ong và chanh: Trộn một thìa mật ong với một thìa nước cốt chanh và uống trước khi đi ngủ. Mật ong có tác dụng làm dịu, trong khi chanh có chứa chất chống oxy hóa có thể giúp cải thiện chất lượng giấc ngủ.

2. Trà thảo dược: Một số loại trà thảo dược như hoa cúc, hoa oải hương và hoa sen có tác dụng thư giãn và giúp bạn chìm vào giấc ngủ. Hãy thử pha một ấm trà thảo dược ấm và uống trước khi đi ngủ.

3. Sữa ấm: Sữa ấm có chứa tryptophan, một loại hormone giúp bạn cảm thấy dễ chịu và thư giãn. Nó cũng có thể giúp thúc đẩy sản xuất melatonin, hormone giúp bạn chìm vào giấc ngủ. Hãy thử uống một cốc sữa ấm trước khi đi ngủ để giúp bạn ngủ ngon hơn.

```
The translated data has been thoroughly cleaned from the coding task, so in theory, the model should be able to code normally by incorporating the coding capabilities from the English coding examples task.
(The translation will translate code definitions and variables.)

I would not recommend running any code generated from this model.
````
> Viết cho tôi một đoạn code bằng python xây dựng một ứng dụng bằng Flask hiển thị 'Xin chào thế giới' trên màn hình

Để viết một đoạn mã Python xây dựng một ứng dụng bằng Flask hiển thị 'Xin chào thế giới' trên màn hình, bạn có thể làm theo các bước sau:

1. Cài đặt Flask và các thư viện cần thiết:

  ```
  pip install flask
  ```

2. Tạo một ứng dụng Flask mới:

  ```
  app = Flask(__name__)
  ```

3. Thêm một route để hiển thị 'Xin chào thế giới':

  ```
  @app.route('/')
  def hello():
     return 'Xin chào thế giới!'
  ```

4. Khởi tạo ứng dụng Flask:

  ```
  app.run()
  ```

5. Chạy ứng dụng Flask:

  ```
  app.run(debug=True)
  ```

6. Mở trình duyệt và truy cập vào địa chỉ IP hoặc tên máy chủ của máy tính hoặc máy chủ nơi bạn đã khởi động ứng dụng Flask.

````
These examples are sub-optimal as they were generated with greedy decoding to ensure reproducibility.
Here's the recommended configuration:
* top_k=50
* top_p=0.5
* temperature=0.5
* penalty_alpha=0.7
* repetition_penalty=1.15

Note: 120GB of pre-trained Vietnamese  data might not be enough for a general question about Vietnamese events.

Here is a kaggle script to quickly test the model:
* https://www.kaggle.com/code/tuanphamm/t-llama-test

## Training Details

**Hardware Type:**
  * GPU: VGA NVIDIA Tesla P100 16GB
  * SYSTEM RAM: 29GB
  
**Hours used:** ~47.5 days Approx*

### Training Data

* BactrianX 
* OpenOrca_translated 
* WizardLM_70k_translated 
* TigerLabMathInstruct_translated_vi 
* GradeSchoolMathInstruct_translated 
* vilm_lima-vi
* MTEngVietnamese 
* databricks_dolly15k_translated 
* AlpacaCleaned_translated 
* databricks_dolly15k
* OpenOrca
* GradeSchoolMathInstruct 
* AlpacaCleaned
* WebglmQA

### Training Procedure

<!-- This relates heavily to the Technical Specifications. Content here should link to that section when it is relevant to the training procedure. -->

* Learning rate: 2e-5 cosine
* Optimizer: PagedLion8bit
* QLora: rank: 64 /Q: 4-bit
* Batch size: 2
* Gradient accumulation: 128
* Effective batch size: 256
  
  - 250k examples of 70% Vietnamese 30% English for 3.37 epoch
  - 350k examples of 60% Vietnamese 40% English for 1.4 epoch

### Training loss

Last 10 runs:

![image/png](https://cdn-uploads.huggingface.co/production/uploads/63905e87df447b438817b2cd/rV8Go_YFZv7QcR_FhFxp-.png)

Each line is 12 hours

All run plots:

![image/png](https://cdn-uploads.huggingface.co/production/uploads/63905e87df447b438817b2cd/bxdq_0Mu8seBJCJNmpoEb.png)

A bad way to visualize i know...

## Evaluation

<!-- This section describes the evaluation protocols and provides the results. -->

![image/png](https://cdn-uploads.huggingface.co/production/uploads/63905e87df447b438817b2cd/z1ZTm7Tab4tQbVPgQW1hU.png)

Our model currently sits at TOP-5 on the VMLU benchmark

## Citation

<!-- If there is a paper or blog post introducing the model, the APA and Bibtex information for that should go in this section. -->
```bibtex
@online{t-llama,
  author = {Pham Minh Tuan},
  title = {T-Llama: A New Language Model for Vietnamese}, % Placeholder
  year = 2024,
  url = {https://github.com/vTuanpham/Vietnamese_QA_System}
}
```