File size: 10,455 Bytes
598baca
 
 
 
 
 
fd50dea
 
 
 
e1f3912
fd50dea
e1f3912
fd50dea
e1f3912
 
 
 
 
fd50dea
e1f3912
fd50dea
e1f3912
fd50dea
 
 
 
 
 
7923492
fd50dea
 
 
 
 
 
 
 
 
 
 
e1f3912
fd50dea
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
e1f3912
fd50dea
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
e1f3912
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
---
license: llama3
language:
- en
- fa
pipeline_tag: text-generation
---

# Model Details

This Repository is a 4-bit quantized version of [Dorna-Llama3-8B-Instruct](https://huggingface.co/PartAI/Dorna-Llama3-8B-Instruct) model for efficient memory usage. Dorna model is a decoder-only model, specifically trained/fine-tuned on Persian data. [Flash Attention 2](https://arxiv.org/abs/2307.08691) is also integrated for faster inference.

## Benefits

- **Reduced Memory Usage**: 4-bit quantization lowers memory requirements.
- **Faster Inference**: Flash Attention 2 speeds up processing.
- **Easy Deployment**: No need for additional libraries like LlamaCPP or Candle.
- **Ready to Use**: Compatible with Langchain, Haystack, LlamaIndex 2, and more.
- **Google Colab Friendly**: Can run on Google Colab free tier with T4 GPU (less than 15 GB of GPU RAM).

## How to use

You can run conversational inference using the Transformers Auto classes with the `generate()` function. Let's look at an example.

```Python
import torch
import transformers
from transformers import AutoTokenizer, AutoModelForCausalLM

model_path = "amirMohammadi/Dorna-Llama3-8B-Instruct-Quantized4Bit"

tokenizer = AutoTokenizer.from_pretrained(model_path)
model = AutoModelForCausalLM.from_pretrained(
    model_path,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)

messages = [
    {"role": "system",
     "content": "You are a helpful Persian assistant. Please answer questions in the asked language."},
    {"role": "user", "content": "اصفهان بزرگ تر است یا قم؟"},
]

input_ids = tokenizer.apply_chat_template(
    messages,
    add_generation_prompt=True,
    return_tensors="pt"
).to(model.device)

terminators = [
    tokenizer.eos_token_id,
    tokenizer.convert_tokens_to_ids("<|eot_id|>")
]

outputs = model.generate(
    input_ids,
    max_new_tokens=256,
    eos_token_id=terminators,
    do_sample=True,
    temperature=0.6,
    top_p=0.9,
)
response = outputs[0][input_ids.shape[-1]:]
print(tokenizer.decode(response, skip_special_tokens=True))
```

## Evaluation of Non-Quantized version

This model is evaluated on questions across various tasks, including Boolean Questions, Code Generation, Long Response, Math, News QA, Paraphrasing, General Knowledge, and Summarization. Most categories typically have two main difficulty levels: Hard and Easy.

Both human evaluation and automatic evaluation (with GPT-4 as the judge) are performed.

In both tables, **Dorna-8B-it** is used as an abbreviated form of **Dorna-Llama3-8B-Instruct**.

Overall human evaluation results are as follows:


|**Model Pairs**            | **Parameters** |**Win %**|**Lose %**|**Tie %**|
|--------------------------|:---------:|:---------:|:---------:|:---------:|
| Dorna-8B-it **vs.** Meta-Llama-3-8B-Instruct  | 8B |**36.94**|  17.39  |  45.67  |
| Dorna-8B-it **vs.** GPT 3.5 turbo-1106        | N.A. |**32.01**|  26.94  |  41.05  |
| Dorna-8B-it **vs.** Persian Mind              | 7B |**55.77**|  10.49  |  33.74  |


Category-based human evaluation results are as follows:

Win/Lose/Tie % is reported for each category.

<!-- | **Model Pairs**                              | **Parameters** | **Bool Complex** | **Bool Easy**  | **Code Gen** | **General Long Response** | **Historical Long Response** | **Math Complex** | **Math Easy**  | **News QA Complex** | **News QA Easy** | **Paraphrasing** | **General Knowledge Easy** | **General Knowledge Hard** | **Summarization** |
|:----------------------------------------------|:------------:|:----------------:|:----------------:|:-------------:|:-----------------------:|:--------------------------:|:----------------:|:----------------:|:-----------------:|:----------------:|:---------------:|:------------------------:|:------------------------:|:---------------:|
| Dorna-8B-it **vs.** Meta-Llama-3-8B-Instruct | 8B         | 0.25/0.25/0.5  | 0.28/0.35/0.38 | 0.6/0.1/0.3 | 0.8/0.08/0.12         | 0.4/0.3/0.3              | 0.28/0.08/0.65 | 0.47/0.00/0.53 | 0.55/0.07/0.38  | 0.43/0.15/0.42 | 0.1/0.05/0.85 | 0.31/0.2/0.49          | 0.59/0.13/0.28         | 0.28/0.2/0.53 |
| Dorna-8B-it **vs.** GPT 3.5 turbo-1106       | N.A.       | 0.35/0.35/0.3  | 0.3/0.3/0.4    | 0.1/0.3/.06 | 0.2/0.45/0.35         | 0.46/0.27/0.27           | 0.25/0.1/0.65  | 0.05/0.1/0.85  | 0.12/0.35/0.53  | 0.15/0.1/0.75  | 0.25/0.15/0.6 | 0.3/0.32/0.38          | 0.22/0.53/0.25         | 0.35/0.55/0.1 |
| Dorna-8B-it **vs.** Persian Mind             | 7B         | 0.47/0.25/0.28 | 0.57/0.15/0.28 | 0.9/0.1/0.0 | 0.82/0.08/0.1         | 0.4/0.17/0.42            | 0.3/0.0/0.7    | 0.22/0.08/0.7  | 0.72/0.07/0.2   | 0.7/0.0/0.3    | 0.7/0.05/0.25 | 0.51/0.12/0.37         | 0.61/0.1/0.29          | 0.93/0.0/0.07 |
-->

<div style="overflow-x: auto;">
  <table>
    <thead>
      <tr style="vertical-align: middle;">
        <th style="white-space: nowrap; vertical-align: middle;"><strong>Model Pairs</strong></th>
        <th style="white-space: nowrap; vertical-align: middle;"><strong>Parameters</strong></th>
        <th style="white-space: nowrap; vertical-align: middle;"><strong>Bool Complex</strong></th>
        <th style="white-space: nowrap; vertical-align: middle;"><strong>Bool Easy</strong></th>
        <th style="white-space: nowrap; vertical-align: middle;"><strong>Code Gen</strong></th>
        <th style="white-space: nowrap; vertical-align: middle;"><strong>General Long Response</strong></th>
        <th style="white-space: nowrap; vertical-align: middle;"><strong>Historical Long Response</strong></th>
        <th style="white-space: nowrap; vertical-align: middle;"><strong>Math Complex</strong></th>
        <th style="white-space: nowrap; vertical-align: middle;"><strong>Math Easy</strong></th>
        <th style="white-space: nowrap; vertical-align: middle;"><strong>News QA Complex</strong></th>
        <th style="white-space: nowrap; vertical-align: middle;"><strong>News QA Easy</strong></th>
        <th style="white-space: nowrap; vertical-align: middle;"><strong>Paraphrasing</strong></th>
        <th style="white-space: nowrap; vertical-align: middle;"><strong>General Knowledge Easy</strong></th>
        <th style="white-space: nowrap; vertical-align: middle;"><strong>General Knowledge Hard</strong></th>
        <th style="white-space: nowrap; vertical-align: middle;"><strong>Summarization</strong></th>
      </tr>
    </thead>
    <tbody>
      <tr>
        <td style="white-space: nowrap; vertical-align: middle;">Dorna-8B-it <strong>vs.</strong> Meta-Llama-3-8B-Instruct</td>
        <td>8B</td>
        <td>0.25/0.25/0.5</td>
        <td>0.28/<strong>0.35</strong>/0.38</td>
        <td><strong>0.6</strong>/0.1/0.3</td>
        <td><strong>0.8</strong>/0.08/0.12</td>
        <td><strong>0.4</strong>/0.3/0.3</td>
        <td><strong>0.28</strong>/0.08/0.65</td>
        <td><strong>0.47</strong>/0.00/0.53</td>
        <td><strong>0.55</strong>/0.07/0.38</td>
        <td><strong>0.43</strong>/0.15/0.42</td>
        <td><strong>0.1</strong>/0.05/0.85</td>
        <td><strong>0.31</strong>/0.2/0.49</td>
        <td><strong>0.59</strong>/0.13/0.28</td>
        <td><strong>0.28</strong>/0.2/0.53</td>
      </tr>
      <tr>
        <td style="white-space: nowrap; vertical-align: middle;">Dorna-8B-it <strong>vs.</strong> GPT 3.5 turbo-1106</td>
        <td>N.A.</td>
        <td>0.35/0.35/0.3</td>
        <td>0.3/0.3/0.4</td>
        <td>0.1/<strong>0.3</strong>/.06</td>
        <td>0.2/<strong>0.45</strong>/0.35</td>
        <td><strong>0.46</strong>/0.27/0.27</td>
        <td><strong>0.25</strong>/0.1/0.65</td>
        <td>0.05/<strong>0.1</strong>/0.85</td>
        <td>0.12/<strong>0.35</strong>/0.53</td>
        <td><strong>0.15</strong>/0.1/0.75</td>
        <td><strong>0.25</strong>/0.15/0.6</td>
        <td>0.3/<strong>0.32</strong>/0.38</td>
        <td>0.22/<strong>0.53</strong>/0.25</td>
        <td>0.35/<strong>0.55</strong>/0.1</td>
      </tr>
      <tr>
        <td style="white-space: nowrap; vertical-align: middle;">Dorna-8B-it <strong>vs.</strong> Persian Mind</td>
        <td>7B</td>
        <td><strong>0.47</strong>/0.25/0.28</td>
        <td><strong>0.57</strong>/0.15/0.28</td>
        <td><strong>0.9</strong>/0.1/0.0</td>
        <td><strong>0.82</strong>/0.08/0.1</td>
        <td><strong>0.4</strong>/0.17/0.42</td>
        <td><strong>0.3</strong>/0.0/0.7</td>
        <td><strong>0.22</strong>/0.08/0.7</td>
        <td><strong>0.72</strong>/0.07/0.2</td>
        <td><strong>0.7</strong>/0.0/0.3</td>
        <td><strong>0.7</strong>/0.05/0.25</td>
        <td><strong>0.51</strong>/0.12/0.37</td>
        <td><strong>0.61</strong>/0.1/0.29</td>
        <td><strong>0.93</strong>/0.0/0.07</td>
      </tr>
    </tbody>
  </table>
</div>



Automatic evaluation results are as follows:


| **Model Pairs**                        | **Parameters** | **Overall Win Rate %** | **Easy Win Rate %** | **Hard Win Rate %** |
|----------------------------------------|:--------------:|:----------------------:|:-------------------:|:-------------------:|
| Dorna-8B-it **vs.** Llama 3 base       |       8B       |       **58.96**        |      **56.00**      |      **64.49**      |
| Dorna-8B-it **vs.** Part Mistral       |       7B       |       **77.20**        |      **73.00**      |      **85.05**      |
| Dorna-8B-it **vs.** Persian Mind       |       7B       |       **90.88**        |      **87.50**      |      **97.20**      |
| Dorna-8B-it **vs.** Neuraorca Gemma 7b |       7B       |       **86.32**        |      **86.50**      |      **85.98**      |
| Dorna-8B-it **vs.** Maral 7b           |       7B       |       **97.39**        |      **97.00**      |      **98.13**      |
| Dorna-8B-it **vs.** PersianLlama 7b    |       7B       |       **98.70**        |      **98.00**      |      **100.00**      |
| Dorna-8B-it **vs.** Aya-23-8B          |       8B       |       **52.77**        |      **56.50**      |        45.79        |
| Dorna-8B-it **vs.** Aya-23-35B         |      35B       |         45.93          |      **54.00**      |        30.84        |
| Dorna-8B-it **vs.** Command R          |      35B       |       **58.63**        |      **61.00**      |      **54.21**      |


## Contact us

If you have any questions regarding this model, you can reach us via the [community](https://huggingface.co/amirMohammadi/Dorna-Llama3-8B-Instruct-Quantized4Bit/discussions) on Hugging Face.