jcai / breeze-7b-instruct-v1_0

实用性方面
- Breeze-7B-Base 在原始词汇的基础上增加了 30,000 个繁体中文词素。由于词汇量的增加，在其他条件相同的情况下，Breeze-7B 在繁体中文上的推理速度是 Mistral-7B 和 Llama 7B 的两倍。[见推理性能。]
- Breeze-7B-Instruct 可以直接用于常见任务，如问答、RAG、多轮对话和摘要。

性能方面
- 与台湾-LLM-7B/13B-chat、QWen(1.5)-7B-Chat 和 Yi-6B-Chat 等同等规模的开源同代产品相比，Breeze-7B-Instruct 在繁体中文和英语的基准测试中表现出令人印象深刻的性能。[见聊天模型性能。]

由以下成员（按字母顺序排列）的项目：许湛然、刘昶乐、廖峰挺、许博竣、陈宜昌和导师许大山。

演示

在这里尝试演示 👩‍💻🧑🏻‍💻

特性

Breeze-7B-Base-v1_0
- 将词汇字典的大小从32k扩展到62k，以更好地支持繁体中文
- 8k令牌上下文长度
Breeze-7B-Instruct-v1_0
- 将词汇字典的大小从32k扩展到62k，以更好地支持繁体中文
- 8k令牌上下文长度
- 多轮对话（不针对有害性进行特殊处理）

模型详细信息

Breeze-7B-Base-v1_0
- 微调自： mistralai/Mistral-7B-v0.1
- 模型类型：因果解码器仅 Transformer 语言模型
- 语言：英语和繁体中文（zh-tw）
Breeze-7B-Instruct-v1_0
- 微调自： MediaTek-Research/Breeze-7B-Base-v1_0
- 模型类型：因果解码器仅 Transformer 语言模型
- 语言：英语和繁体中文（zh-tw）

基础模型性能

在这里，我们比较 Breeze-7B-Base-v1_0 与其他广泛认可的类似参数规模的公开基础语言模型，这些模型在中国有非常好的性能。
TMMLU+、DRCD 和 Table 数据来源为 MediaTek-Research/TCEval-v2。
MediaTek-Research/TCEval-v2 来自 TCEval-v1
和 ikala/tmmluplus。 MMLU 来源为 hails/mmlu_no_train。
我们使用从 EleutherAI/lm-evaluation-harness 改进的代码来评估 TMMLU+、DRCD、Table 和 MMLU。所有选择问题都通过对数似然率进行调整。

模型	#参数数量	↑ TMMLU+ (ACC)	DRCD (EM)	Table (ACC)	MMLU (ACC)
		TC, 知识	TC, 推理	TC, 推理	EN, 知识
		5 shot	3 shot	5 shot	5 shot
Yi-6B	6B	49.63	76.61	34.72	65.35
Qwen1.5-7B	7B	46.59	74.41	30.56	63.07
Breeze-7B-Base-v1_0	7B	42.67	80.61	31.99	61.24
Mistral-7B-v0.1	7B	36.93	79.27	27.78	64.89

指令微调模型性能

在这里，我们比较 Breeze-7B-Instruct-v1_0 与其他公开指令微调的语言模型，这些模型在中文中的性能广泛认可。
我们还列出了 GPT-3.5 Turbo (1106) 的基准分数，它代表了一种广泛使用的高质量云语言模型 API 服务，仅供参考。
TMMLU+、DRCD、Table 和 MT-Bench-tw 数据来源为 MediaTek-Research/TCEval-v2。
MediaTek-Research/TCEval-v2 来自 TCEval-v1
和 ikala/tmmluplus。 MMLU 来源为 hails/mmlu_no_train。
MT-Bench 数据来源为 lmsys/mt_bench_human_judgments。
我们使用从 EleutherAI/lm-evaluation-harness 改进的代码来评估 TMMLU+、DRCD、Table 和 MMLU。所有选择问题都通过对数似然率进行调整。
我们使用从 fastchat llm_judge (GPT4 作为评判者) 改进的代码来评估 MT-Bench-tw 和 MT-Bench。

模型	#参数数量	↑ MT-Bench-tw (得分)	TMMLU+ (ACC)	Table (ACC)	MT-Bench (得分)	MMLU (ACC)
		TC, 聊天	TC, 知识	TC, 推理	EN, 聊天	EN, 知识
		0 shot	0 shot	0 shot	0 shot	0 shot
GPT-3.5-Turbo		7.1	43.56	45.14	7.9	67.09
Qwen1.5-7B-Chat	7B	6.4	45.65	34.72	7.6	61.85
Breeze-7B-Instruct-v1_0	7B	6.0	42.67	39.58	7.4	61.73
Mistral-7B-v0.2-Instruct	7B	5.6	34.95	33.33	7.6	59.97
Yi-6B-Chat	6B	5.0	44.79	25.69	6.0	59.45
台湾-LLM-13B-v2.0-chat	13B	5.0	29.47	23.61	N/A*	50.50
台湾-LLM-7B-v2.1-chat	7B	4.2	28.08	31.25	N/A*	42.72

* 台湾-LLM 模型用繁体中文回答（英语）多轮问题。

关于 MT-Bench-tw (0 shot) 的详细信息模型	STEM	提取	推理	数学	编码	角色扮演	写作	人文学科	AVG
GPT-3.5-Turbo	7.8	6.1	5.1	6.4	6.2	8.7	7.4	9.3	7.1
Qwen1.5-7B-Chat	9	5.6	4.7	2.8	3.7	8.0	8.0	9.4	6.4
Breeze-7B-Instruct-v1_0	7.8	5.2	4.2	4.2	4.1	7.6	5.9	9.1	6.0
Mistral-7B-v0.2-Instruct	6.9	4.6	4.3	3.3	4.4	7.2	6.2	7.8	5.6
Yi-6B-Chat	7.3	2.7	3.1	3.3	2.3	7.2	5.2	8.8	5.0
台湾-LLM-13B-v2.0-chat	6.1	3.4	4.1	2.3	3.1	7.4	6.6	6.8	5.0
台湾-LLM-7B-v2.1-chat	5.2	2.6	2.3	1.2	3.4	6.6	5.7	6.8	4.2

关于 TMMLU+ (0 shot) 的详细信息模型	STEM	社会科学	人文学科	其他	AVG
GPT-3.5-Turbo	41.58	48.52	40.96	43.18	43.56
Qwen1.5-7B-Chat	41.48	51.66	44.05	45.40	45.65
Breeze-7B-Instruct-v1_0	36.46	48.38	45.11	40.75	42.67
Mistral-7B-v0.2-Instruct	32.79	38.05	34.89	34.04	34.94
Yi-6B-Chat	37.80	51.74	45.36	44.25	44.79
台湾-LLM-13B-v2.0-chat	27.74	33.69	27.03	29.43	29.47
台湾-LLM-7B-v2.1-chat	25.58	31.76	27.36	27.61	28.08

推理性能

在本测试中，我们使用网络文章的前700个字符作为输入，要求模型再次撰写相同的文章。
所有推理均在2个RTX A6000 GPU上运行（使用 vllm，使用2个张量并行的大小）。

模型	↓ 推理时间（秒）	最大输入长度估计（字符）
Qwen1.5-7B-Chat	9.35	38.9k
Yi-6B-Chat	10.62	5.2k
Breeze-7B-Instruct-v1_0	10.74	11.1k
Mistral-7B-Instruct-v0.2	20.48	5.1k
台湾-LLM-7B-v2.1-chat	26.26	2.2k
<!—	台湾-LLM-13B-v2.0-chat	36.80

在 Transformers 中使用

首先安装直接依赖项

pip install transformers torch accelerate

如果要使用 flash-attention2 快速进行推理，则需要安装这些依赖项

pip install packaging ninja
pip install flash-attn

然后在 transformers 中加载模型

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

# Instruction Model
model = AutoModelForCausalLM.from_pretrained(
    "MediaTek-Research/Breeze-7B-Instruct-v1_0",
    device_map="auto",
    torch_dtype=torch.bfloat16,
    # attn_implementation="flash_attention_2" # optional
)

# Basemodel
model = AutoModelForCausalLM.from_pretrained(
    "MediaTek-Research/Breeze-7B-Base-v1_0",
    device_map="auto",
    torch_dtype=torch.bfloat16,
    # attn_implementation="flash_attention_2" # optional
)

Breeze-7B-Instruct 的查询结构如下：

<s>SYS_PROMPT  [INST] QUERY1 [/INST] RESPONSE1 [INST] QUERY2 [/INST]

其中 SYS_PROMPT，QUERY1，RESPONSE1 和 QUERY2 可以由用户提供。

建议的默认 SYS_PROMPT 是：

You are a helpful AI assistant built by MediaTek Research. The user you are helping speaks Traditional Chinese and comes from Taiwan.

我们还把 chat_template 整合到了 tokenizer_config.json 中，因此你可以使用 apply_chat_template 来获取提示。

>>> from transformers import AutoTokenizer
>>> tokenizer = AutoTokenizer.from_pretrained("MediaTek-Research/Breeze-7B-Instruct-v1_0")
>>> chat = [
...   {"role": "user", "content": "你好，請問你可以完成什麼任務？"},
...   {"role": "assistant", "content": "你好，我可以幫助您解決各種問題、提供資訊和協助您完成許多不同的任務。例如：回答技術問題、提供建議、翻譯文字、尋找資料或協助您安排行程等。請告訴我如何能幫助您。"},
...   {"role": "user", "content": "太棒了！"},
... ]
>>> tokenizer.apply_chat_template(chat, tokenize=False)
"<s>You are a helpful AI assistant built by MediaTek Research. The user you are helping speaks Traditional Chinese and comes from Taiwan.  [INST] 你好，請問你可以完成什麼任務？ [/INST] 你好，我可以幫助您解決各種問題、提供資訊和協助您完成許多不同的任務。例如：回答技術問題、提供建議、翻譯文字、尋找資料或協助您安排行程等。請告訴我如何能幫助您。 [INST] 太棒了！ [/INST] "
# Tokenized results
# ['▁', '你好', '，', '請問', '你', '可以', '完成', '什麼', '任務', '？']
# ['▁', '你好', '，', '我', '可以', '幫助', '您', '解決', '各種', '問題', '、', '提供', '資訊', '和', '協助', '您', '完成', '許多', '不同', '的', '任務', '。', '例如', '：', '回答', '技術', '問題', '、', '提供', '建議', '、', '翻譯', '文字', '、', '尋找', '資料', '或', '協助', '您', '安排', '行程', '等', '。', '請', '告訴', '我', '如何', '能', '幫助', '您', '。']
# ['▁', '太', '棒', '了', '！']

文本生成可以通过 generate 和 apply_chat_template 函数完成。

>>> outputs = model.generate(tokenizer.apply_chat_template(chat, return_tensors="pt"),
>>>                          # adjust below parameters if necessary 
>>>                          max_new_tokens=128,
>>>                          top_p=0.01,
>>>                          top_k=85,
>>>                          repetition_penalty=1.1,
>>>                          temperature=0.01)
>>>                          
>>> print(tokenizer.decode(outputs[0]))

参考文献

@article{MediaTek-Research2024breeze7b,
      title={Breeze-7B Technical Report}, 
      author={Chan-Jan Hsu and Chang-Le Liu and Feng-Ting Liao and Po-Chun Hsu and Yi-Chang Chen and Da-Shan Shiu},
      year={2024},
      eprint={2403.02712},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}

**Source: [https://hugging-face.cn/MediaTek-Research/Breeze-7B-Instruct-v1_0](https://hugging-face.cn/MediaTek-Research/Breeze-7B-Instruct-v1_0)**

# MediaTek Research Breeze-7B-Instruct-v1_0

MediaTek Research Breeze-7B (hereinafter referred to as Breeze-7B) is a language model family that builds on top of [Mistral-7B](https://hugging-face.cn/mistralai/Mistral-7B-v0.1), specifically intended for Traditional Chinese use.

[Breeze-7B-Base](https://hugging-face.cn/MediaTek-Research/Breeze-7B-Base-v1_0) is the base model for the Breeze-7B series. 
It is suitable for use if you have substantial fine-tuning data to tune it for your specific use case.

[Breeze-7B-Instruct](https://hugging-face.cn/MediaTek-Research/Breeze-7B-Instruct-v1_0) derives from the base model Breeze-7B-Base, making the resulting model amenable to be used as-is for commonly seen tasks.

The current release version of Breeze-7B is v1.0, which has undergone a more refined training process compared to Breeze-7B-v0_1, resulting in significantly improved performance in both English and Traditional Chinese.

For details of this model please read our [paper](https://arxiv.org/abs/2403.02712).

Practicality-wise:
- Breeze-7B-Base expands the original vocabulary with an additional 30,000 Traditional Chinese tokens. With the expanded vocabulary, and everything else being equal, Breeze-7B operates at twice the inference speed for Traditional Chinese to Mistral-7B and Llama 7B. [See [Inference Performance](#inference-performance).]
- Breeze-7B-Instruct can be used as is for common tasks such as Q&A, RAG, multi-round chat, and summarization.

Performance-wise:
- Breeze-7B-Instruct demonstrates impressive performance in benchmarks for Traditional Chinese and English when compared to similar-sized open-source contemporaries such as Taiwan-LLM-7B/13B-chat, QWen(1.5)-7B-Chat, and Yi-6B-Chat. [See [Chat Model Performance](#chat-model-performance).]

*A project by the members (in alphabetical order): Chan-Jan Hsu 許湛然, Chang-Le Liu 劉昶樂, Feng-Ting Liao 廖峰挺, Po-Chun Hsu 許博竣, Yi-Chang Chen 陳宜昌, and the supervisor Da-Shan Shiu 許大山.*

## Demo

## Features

- Breeze-7B-Base-v1_0
  - Expanding the vocabulary dictionary size from 32k to 62k to better support Traditional Chinese
  - 8k-token context length
- Breeze-7B-Instruct-v1_0
  - Expanding the vocabulary dictionary size from 32k to 62k to better support Traditional Chinese 
  - 8k-token context length
  - Multi-turn dialogue (without special handling for harmfulness)

## Model Details

- Breeze-7B-Base-v1_0
  - Finetuned from: [mistralai/Mistral-7B-v0.1](https://hugging-face.cn/mistralai/Mistral-7B-v0.1)
  - Model type: Causal decoder-only transformer language model
  - Language: English and Traditional Chinese (zh-tw)
- Breeze-7B-Instruct-v1_0
  - Finetuned from: [MediaTek-Research/Breeze-7B-Base-v1_0](https://hugging-face.cn/MediaTek-Research/Breeze-7B-Base-v1_0)
  - Model type: Causal decoder-only transformer language model
  - Language: English and Traditional Chinese (zh-tw)

## Base Model Performance

Here we compare Breeze-7B-Base-v1_0 with other open-source base language models of similar parameter size that are widely recognized for their good performance in Chinese.
**TMMLU+**, **DRCD**, and **Table** source from [MediaTek-Research/TCEval-v2](https://hugging-face.cn/datasets/MediaTek-Research/TCEval-v2).
[MediaTek-Research/TCEval-v2](https://hugging-face.cn/datasets/MediaTek-Research/TCEval-v2) derives from [TCEval-v1](https://github.com/mtkresearch/MR-Models/tree/main/TC-Eval)
 and [ikala/tmmluplus](https://hugging-face.cn/datasets/ikala/tmmluplus). **MMLU** sources from [hails/mmlu_no_train](https://hugging-face.cn/datasets/hails/mmlu_no_train).
 We use the code revised from [EleutherAI/lm-evaluation-harness](https://github.com/EleutherAI/lm-evaluation-harness) to evaluate **TMMLU+**, **DRCD**, **Table**, and **MMLU**. All choice problems adapt the selection by the log-likelihood.

| Models                                                                                    | #Parameters | ↑ TMMLU+ (ACC) | DRCD (EM)   | Table (ACC) | MMLU (ACC) |
|----------------------------------------------                                             |--------|--------------|-------------|-------------|------------|
|                                                                                           |        |TC, Knowledge |TC, Reasoning|TC, Reasoning|EN, Knowledge|
|                                                                                           |        | 5 shot       | 3 shot      | 5 shot      | 5 shot      |
| [Yi-6B](https://hugging-face.cn/01-ai/Yi-6B)                                               | 6B     |   49.63      | 76.61       |   34.72     | 65.35       |
| [Qwen1.5-7B](https://hugging-face.cn/Qwen/Qwen1.5-7B)                                      | 7B     |   46.59      | 74.41       |   30.56     | 63.07       |
| [**Breeze-7B-Base-v1_0**](https://hugging-face.cn/MediaTek-Research/Breeze-7B-Base-v1_0)   | 7B     |   42.67      | 80.61       |   31.99     | 61.24       |
| [Mistral-7B-v0.1](https://hugging-face.cn/mistralai/Mistral-7B-v0.1)                       | 7B     |   36.93      | 79.27       |   27.78     | 64.89       |

## Instruction-tuned Model Performance

Here we compare Breeze-7B-Instruct-v1_0 with other open-source instruction-tuned language models of similar parameter size that are widely recognized for their good performance in Chinese.
Also, we listed the benchmark scores of GPT-3.5 Turbo (1106), which represents one of the most widely used high-quality cloud language model API services, for reference.
**TMMLU+**, **DRCD**, **Table**, and **MT-Bench-tw** source from [MediaTek-Research/TCEval-v2](https://hugging-face.cn/datasets/MediaTek-Research/TCEval-v2).
[MediaTek-Research/TCEval-v2](https://hugging-face.cn/datasets/MediaTek-Research/TCEval-v2) derives from [TCEval-v1](https://github.com/mtkresearch/MR-Models/tree/main/TC-Eval)
 and [ikala/tmmluplus](https://hugging-face.cn/datasets/ikala/tmmluplus). **MMLU** sources from [hails/mmlu_no_train](https://hugging-face.cn/datasets/hails/mmlu_no_train).
 **MT-Bench** source from [lmsys/mt_bench_human_judgments](https://hugging-face.cn/datasets/lmsys/mt_bench_human_judgments).
 We use the code revised from [EleutherAI/lm-evaluation-harness](https://github.com/EleutherAI/lm-evaluation-harness) to evaluate **TMMLU+**, **DRCD**, **Table**, and **MMLU**. All choice problems adapt the selection by the log-likelihood.
 We use the code revised from [fastchat llm_judge](https://github.com/lm-sys/FastChat/tree/main/fastchat/llm_judge) (GPT4 as judge) to evaluate **MT-Bench-tw** and **MT-Bench**.

| Models                                                                                             | #Parameters | ↑ MT-Bench-tw (Score)| TMMLU+ (ACC) | Table (ACC) | MT-Bench (Score) | MMLU (ACC)  | 
|---------------------------------------------------------------------------------------------------------|--------|--------------------|--------------|-------------|------------------|-------------|
|                                                                                                         |        |TC, Chat            |TC, Knowledge |TC, Reasoning|EN, Chat          |EN, Knowledge|
|                                                                                                         |        |0 shot              | 0 shot       | 0 shot      |0 shot            |  0 shot     | 
| [GPT-3.5-Turbo](https://openai.com)                                                                     |        |7.1                 | 43.56        | 45.14       |7.9               |  67.09      |    
| [Qwen1.5-7B-Chat](https://hugging-face.cn/Qwen/Qwen1.5-7B-Chat)                                          | 7B     |6.4                 | 45.65        | 34.72       |7.6               |  61.85      |    
| [**Breeze-7B-Instruct-v1_0**](https://hugging-face.cn/MediaTek-Research/Breeze-7B-Instruct-v1_0)         | 7B     |6.0                 | 42.67        | 39.58       |7.4               |  61.73     |    
| [Mistral-7B-v0.2-Instruct](https://hugging-face.cn/mistralai/Mistral-7B-Instruct-v0.2)                   | 7B     |5.6                 | 34.95        | 33.33       |7.6               |    59.97    |                                                  
| [Yi-6B-Chat](https://hugging-face.cn/01-ai/Yi-6B-Chat)                                                   | 6B     |5.0                 | 44.79        | 25.69       |6.0               |    59.45    |    
| [Taiwan-LLM-13B-v2.0-chat](https://hugging-face.cn/yentinglin/Taiwan-LLM-13B-v2.0-chat)                  | 13B    |5.0                 | 29.47        | 23.61       |N/A*                |    50.50    |     
| [Taiwan-LLM-7B-v2.1-chat](https://hugging-face.cn/yentinglin/Taiwan-LLM-7B-v2.1-chat)                    | 7B     |4.2                 | 28.08        | 31.25       |N/A*               |    42.72    |

\* Taiwan-LLM models respond to multi-turn questions (English) in Traditional Chinese.

| Details on MT-Bench-tw (0 shot):<br/>Models         | STEM    |Extraction|Reasoning| Math   | Coding  | Roleplay| Writing |Humanities|       AVG    | 
|-----------------------------------------------------|---------|---------|---------|---------|---------|---------|---------|----------|  ---------   | 
| GPT-3.5-Turbo                                       |  7.8    |  6.1    |   5.1   |   6.4   |  6.2    |   8.7   |   7.4   |   9.3    |        7.1   |
| Qwen1.5-7B-Chat                                     |  9      |  5.6    |   4.7   |   2.8   |  3.7    |   8.0   |   8.0   |   9.4    |        6.4   |
| **Breeze-7B-Instruct-v1_0**                         |  7.8    |  5.2    |   4.2   |   4.2   |  4.1    |   7.6   |   5.9   |   9.1    |        6.0   |
| Mistral-7B-v0.2-Instruct                            |  6.9    |  4.6    |   4.3   |   3.3   |  4.4    |   7.2   |   6.2   |   7.8    |        5.6   |                                          
| Yi-6B-Chat                                          |  7.3    |  2.7    |   3.1   |   3.3   |  2.3    |   7.2   |   5.2   |   8.8    |        5.0   |
| Taiwan-LLM-13B-v2.0-chat                            |  6.1    |  3.4    |   4.1   |   2.3   |  3.1    |   7.4   |   6.6   |   6.8    |        5.0   |
| Taiwan-LLM-7B-v2.1-chat                             |  5.2    |  2.6    |   2.3   |   1.2   |  3.4    |   6.6   |   5.7   |   6.8    |        4.2   |

| Details on TMMLU+ (0 shot):<br/>Model               | STEM         | Social Science | Humanities | Other      |   AVG   |
|-----------------------------------------------------|--------------|----------------|------------|------------|---------|
| GPT-3.5-Turbo                                       | 41.58        | 48.52          | 40.96      | 43.18      | 43.56   |
| Qwen1.5-7B-Chat                                     | 41.48        | 51.66          | 44.05      | 45.40      | 45.65   |
| **Breeze-7B-Instruct-v1_0**                         | 36.46        | 48.38          | 45.11      | 40.75      | 42.67   |
| Mistral-7B-v0.2-Instruct                            | 32.79        | 38.05          | 34.89      | 34.04      | 34.94   |
| Yi-6B-Chat                                          | 37.80        | 51.74          | 45.36      | 44.25      | 44.79   |
| Taiwan-LLM-13B-v2.0-chat                            | 27.74        | 33.69          | 27.03      | 29.43      | 29.47   |
| Taiwan-LLM-7B-v2.1-chat                             | 25.58        | 31.76          | 27.36      | 27.61      | 28.08   |

## Inference Performance
In this test, we use the first 700 characters of the [web article](https://health.udn.com/health/story/5976/7699252?from=udn_ch1005_main_index) as the input and ask the model to write the same article again.
All inferences run on 2 RTX A6000 GPUs (using `vllm`, with a tensor-parallel size of 2).

| Models                                                             | ↓ Inference Time (sec)|Estimated Max Input Length (Char)|
|--------------------------------------------------------------------|-------------------|--------------------------|
| Qwen1.5-7B-Chat                                                    |  9.35             |    38.9k                  |
| Yi-6B-Chat                                                         |   10.62           |   5.2k                |
| **Breeze-7B-Instruct-v1_0**                                        |  10.74            |    11.1k                 |
| Mistral-7B-Instruct-v0.2                                           |  20.48           |    5.1k                 |
| Taiwan-LLM-7B-v2.1-chat                                            |   26.26          |    2.2k                  |

## Use in Transformers

First install direct dependencies:
```
pip install transformers torch accelerate
```
If you want faster inference using flash-attention2, you need to install these dependencies:
```bash
pip install packaging ninja
pip install flash-attn
```
Then load the model in transformers:
```python
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

# Instruction Model
model = AutoModelForCausalLM.from_pretrained(
    "MediaTek-Research/Breeze-7B-Instruct-v1_0",
    device_map="auto",
    torch_dtype=torch.bfloat16,
    # attn_implementation="flash_attention_2" # optional
)

# Basemodel
model = AutoModelForCausalLM.from_pretrained(
    "MediaTek-Research/Breeze-7B-Base-v1_0",
    device_map="auto",
    torch_dtype=torch.bfloat16,
    # attn_implementation="flash_attention_2" # optional
)
```

**For Breeze-7B-Instruct**, the structure of the query is 
```txt
<s>SYS_PROMPT  [INST] QUERY1 [/INST] RESPONSE1 [INST] QUERY2 [/INST] 
```
where `SYS_PROMPT`, `QUERY1`, `RESPONSE1`, and `QUERY2` can be provided by the user.

The suggested default `SYS_PROMPT` is 
```txt
You are a helpful AI assistant built by MediaTek Research. The user you are helping speaks Traditional Chinese and comes from Taiwan.
```

We also integrate `chat_template` into [tokenizer_config.json](tokenizer_config.json), so you can `apply_chat_template` to get the prompt.

```python
>>> from transformers import AutoTokenizer
>>> tokenizer = AutoTokenizer.from_pretrained("MediaTek-Research/Breeze-7B-Instruct-v1_0")
>>> chat = [
...   {"role": "user", "content": "你好，請問你可以完成什麼任務？"},
...   {"role": "assistant", "content": "你好，我可以幫助您解決各種問題、提供資訊和協助您完成許多不同的任務。例如：回答技術問題、提供建議、翻譯文字、尋找資料或協助您安排行程等。請告訴我如何能幫助您。"},
...   {"role": "user", "content": "太棒了！"},
... ]
>>> tokenizer.apply_chat_template(chat, tokenize=False)
"<s>You are a helpful AI assistant built by MediaTek Research. The user you are helping speaks Traditional Chinese and comes from Taiwan.  [INST] 你好，請問你可以完成什麼任務？ [/INST] 你好，我可以幫助您解決各種問題、提供資訊和協助您完成許多不同的任務。例如：回答技術問題、提供建議、翻譯文字、尋找資料或協助您安排行程等。請告訴我如何能幫助您。 [INST] 太棒了！ [/INST] "
# Tokenized results
# ['▁', '你好', '，', '請問', '你', '可以', '完成', '什麼', '任務', '？']
# ['▁', '你好', '，', '我', '可以', '幫助', '您', '解決', '各種', '問題', '、', '提供', '資訊', '和', '協助', '您', '完成', '許多', '不同', '的', '任務', '。', '例如', '：', '回答', '技術', '問題', '、', '提供', '建議', '、', '翻譯', '文字', '、', '尋找', '資料', '或', '協助', '您', '安排', '行程', '等', '。', '請', '告訴', '我', '如何', '能', '幫助', '您', '。']
# ['▁', '太', '棒', '了', '！']
```

Text generation can be done by `generate` and `apply_chat_template` functions:
```python
>>> outputs = model.generate(tokenizer.apply_chat_template(chat, return_tensors="pt"),
>>>                          # adjust below parameters if necessary 
>>>                          max_new_tokens=128,
>>>                          top_p=0.01,
>>>                          top_k=85,
>>>                          repetition_penalty=1.1,
>>>                          temperature=0.01)
>>>                          
>>> print(tokenizer.decode(outputs[0]))
```

## Citation

```
@article{MediaTek-Research2024breeze7b,
      title={Breeze-7B Technical Report}, 
      author={Chan-Jan Hsu and Chang-Le Liu and Feng-Ting Liao and Po-Chun Hsu and Yi-Chang Chen and Da-Shan Shiu},
      year={2024},
      eprint={2403.02712},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}
```

Paste, drop or click to upload images (.png, .jpeg, .jpg, .svg, .gif)