File size: 4,393 Bytes
282c302
 
 
 
30e97d7
282c302
 
d02976b
 
327aac8
d02976b
 
 
 
 
 
 
 
 
 
 
 
b911773
d02976b
 
 
 
 
 
 
 
 
 
 
327aac8
f4c1548
d02976b
327aac8
f4c1548
d02976b
327aac8
f4c1548
d02976b
327aac8
f4c1548
d02976b
327aac8
f4c1548
d02976b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1dbdba8
d02976b
 
 
 
da0c039
 
d02976b
 
 
 
 
 
f4c1548
a885f37
d02976b
 
 
 
 
 
 
32c8482
f4c1548
d02976b
 
 
 
 
 
 
 
 
30e97d7
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
---
license: apache-2.0
language:
- zh
pipeline_tag: text-generation
---

# Hyacinth6B: A Trandidional Chinese Large Language Model 

<img src="./pics/hyacinth.jpeg" alt="image_name png"/>

Hyacinth6B is a Tranditional Chinese Large Language Model which fine-tune from [chatglm3-base](https://huggingface.co/THUDM/chatglm3-6b-base),our goal is to find a balance between model lightness and performance, striving to maximize performance while using a comparatively lightweight model. Hyacinth6B was developed with this objective in mind, aiming to fully leverage the core capabilities of LLMs without incurring substantial resource costs, effectively pushing the boundaries of smaller models' performance. The training approach involves parameter-efficient fine-tuning using the Low-Rank Adaptation (LoRA) method.
At last, we evaluated Hyacinth6B, examining its performance across various aspects. Hyacinth6B shows commendable performance in certain metrics, even surpassing ChatGPT in two categories. We look forward to providing more resources and possibilities for the field of Traditional Chinese language processing. This research aims to expand the research scope of Traditional Chinese language models and enhance their applicability in different scenarios.








# Training Config
Training required approximately 20.6GB of VRAM without any quantization (default fp16) and a total of 369 hours in duration on single RTX 4090.

| HyperParameter | Value |
| --------- | ----- |
| Batch Size|   8   |
|Learning Rate |5e-5       |
|Epochs    |3       |
|LoRA r| 16  |


# Evaluate Results
## CMMLU
<img src="./pics/cmmlu.png" alt="image_name png"/>

## C-eval
<img src="./pics/ceval.png" alt="image_name png"/>

## TC-eval by MediaTek Research
<img src="./pics/tc-eval.png" alt="image_name png"/>

## MT-bench
<img src="./pics/dashB.png" alt="image_name png"/>

## LLM-eval by NTU Miu Lab
<img src="./pics/llmeval.png" alt="image_name png"/>

## Bailong Bench


| Bailong-bench| Taiwan-LLM-7B-v2.1-chat |Taiwan-LLM-13B-v2.0-chat |gpt-3.5-turbo-1103|Bailong-instruct 7B|Hyacinth6B(ours)|
| -------- | -------- | --- | --- | --- | -------- |
|Arithmetic|9.0|10.0|10.0|9.2|8.4|
|Copywriting generation|7.6|3.0|9.0|9.6|10.0 |
|Creative writing|6.1|7.5     |8.7     |9.4     |8.3 |
|English instruction| 6.0| 1.9 |10.0 |9.2  |  10.0  |
|General|7.7| 8.1    |9.9     |9.2     |         9.2 |
|Health consultation|7.7|  8.5   |9.9     |9.2     |  9.8  |
|Knowledge-based question|4.2| 8.4    | 9.9    | 9.8    |4.9 |
|Mail assistant|9.5|  9.9   |9.0     |9.9     |    9.5      |
|Morality and Ethics|   4.5       | 9.3    |9.8     |9.7     |7.4 |
|Multi-turn|7.9|8.7     |9.0     |7.8     |4.4          |
|Open question|7.0|9.2     |7.6     |9.6     | 8.2         |
|Proofreading|3.0|4.0     |10.0     |9.0     |       9.1   |
|Summarization|6.2| 7.4    |9.9     |9.8     |     8.4     |
|Translation|7.0|9.0     |8.1     |9.5     |   10.0       |
|**Average**|6.7| 7.9    |9.4     |9.4     |  8.4    |


## Acknowledgement
Thanks for Taiwan LLM's author, Yen-Ting Lin 's kindly advice.
Please review his marvellous works!
[Yen-Ting Lin's hugging face](https://huggingface.co/yentinglin)


## Disclaimer
This model is intended for research purposes only. The author does not guarantee its accuracy, completeness, or suitability for any purpose. Any commercial or other use requires consultation with a legal professional, and the author assumes no responsibility for such use. Users bear all risks associated with the results of using this model. The author is not liable for any direct or indirect losses or damages, including but not limited to loss of profits, business interruption, or data loss. Any use of this model is considered acceptance of the terms of this disclaimer.

### Model Usage

Download model

Here is the example for you to download Hyacinth6B with huggingface transformers:

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

tokenizer = AutoTokenizer.from_pretrained("chillymiao/Hyacinth6B")
model = AutoModelForCausalLM.from_pretrained("chillymiao/Hyacinth6B")
```

### Citation

```
@misc{song2024hyacinth6b,
      title={Hyacinth6B: A large language model for Traditional Chinese}, 
      author={Chih-Wei Song and Yin-Te Tsai},
      year={2024},
      eprint={2403.13334},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}
```