Hugging Face's logo

RichardErkhov
/
MediaTek-Research_-_Breeze-7B-Instruct-v0_1-4bits

Text Generation
Transformers
Safetensors
mistral
conversational
text-generation-inference
4-bit precision
bitsandbytes
Model card Files Community
RichardErkhov commited on
Commit
d0f268c
·
verified ·
1 Parent(s): d0b6922

uploaded readme

Browse files
Files changed (1) hide show
  1. README.md +247 -0
README.md ADDED
@@ -0,0 +1,247 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ Quantization made by Richard Erkhov.
2
+
3
+ [Github](https://github.com/RichardErkhov)
4
+
5
+ [Discord](https://discord.gg/pvy7H8DZMG)
6
+
7
+ [Request more models](https://github.com/RichardErkhov/quant_request)
8
+
9
+
10
+ Breeze-7B-Instruct-v0_1 - bnb 4bits
11
+ - Model creator: https://huggingface.co/MediaTek-Research/
12
+ - Original model: https://huggingface.co/MediaTek-Research/Breeze-7B-Instruct-v0_1/
13
+
14
+
15
+
16
+
17
+ Original model description:
18
+ ---
19
+ pipeline_tag: text-generation
20
+ license: apache-2.0
21
+ language:
22
+ - zh
23
+ - en
24
+ ---
25
+
26
+ # Model Card for MediaTek Research Breeze-7B-Instruct-v0_1
27
+
28
+ MediaTek Research Breeze-7B (hereinafter referred to as Breeze-7B) is a language model family that builds on top of [Mistral-7B](https://huggingface.co/mistralai/Mistral-7B-v0.1), specifically intended for Traditional Chinese use.
29
+
30
+ [Breeze-7B-Base](https://huggingface.co/MediaTek-Research/Breeze-7B-Base-v0_1) is the base model for the Breeze-7B series.
31
+ It is suitable for use if you have substantial fine-tuning data to tune it for your specific use case.
32
+
33
+ [Breeze-7B-Instruct](https://huggingface.co/MediaTek-Research/Breeze-7B-Instruct-v0_1) derives from the base model Breeze-7B-Base, making the resulting model amenable to be used as-is for commonly seen tasks.
34
+
35
+ [Breeze-7B-Instruct-64k](https://huggingface.co/MediaTek-Research/Breeze-7B-Instruct-64k-v0_1) is a slightly modified version of
36
+ Breeze-7B-Instruct to enable a 64k-token context length. Roughly speaking, that is equivalent to 88k Traditional Chinese characters.
37
+
38
+ *Update (Feb. 21st, 2024): Breeze-7B-Instruct-64k-v0_1 has been temporarily removed from Hugging Face due to its actual performance in long context tests not meeting expectations.*
39
+
40
+ *Update (Mar. 7th, 2024): The current release version of Breeze-7B is v1.0. See [Breeze-7B-Instruct-v1_0](https://huggingface.co/MediaTek-Research/Breeze-7B-Instruct-v1_0).*
41
+
42
+ The current release version of Breeze-7B is v0.1.
43
+
44
+ Practicality-wise:
45
+ - Breeze-7B-Base expands the original vocabulary with additional 30,000 Traditional Chinese tokens. With the expanded vocabulary, 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).]
46
+ - Breeze-7B-Instruct can be used as is for common tasks such as Q&A, RAG, multi-round chat, and summarization.
47
+ - In particular, Breeze-7B-Instruct-64k can perform tasks at a document level, not a chapter level.
48
+
49
+
50
+ Performance-wise:
51
+ - 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-7B-Chat, and Yi-6B-Chat. [See [Chat Model Performance](#chat-model-performance).]
52
+
53
+
54
+ *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.*
55
+
56
+ ## Features
57
+
58
+ - Breeze-7B-Base-v0_1
59
+ - Expanding the vocabulary dictionary size from 32k to 62k to better support Traditional Chinese
60
+ - 8k-token context length
61
+ - Breeze-7B-Instruct-v0_1
62
+ - Expanding the vocabulary dictionary size from 32k to 62k to better support Traditional Chinese
63
+ - 8k-token context length
64
+ - Multi-turn dialogue (without special handling for harmfulness)
65
+ - Breeze-7B-Instruct-64k-v0_1
66
+ - Expanding the vocabulary dictionary size from 32k to 62k to better support Traditional Chinese
67
+ - 64k-token context length
68
+ - Multi-turn dialogue (without special handling for harmfulness)
69
+
70
+ ## Model Details
71
+
72
+ - Breeze-7B-Base-v0_1
73
+ - Finetuned from: [mistralai/Mistral-7B-v0.1](https://huggingface.co/mistralai/Mistral-7B-v0.1)
74
+ - Model type: Causal decoder-only transformer language model
75
+ - Language: English and Traditional Chinese (zh-tw)
76
+ - Breeze-7B-Instruct-v0_1
77
+ - Finetuned from: [MediaTek-Research/Breeze-7B-Base-v0_1](https://huggingface.co/MediaTek-Research/Breeze-7B-Base-v0_1)
78
+ - Model type: Causal decoder-only transformer language model
79
+ - Language: English and Traditional Chinese (zh-tw)
80
+ - Breeze-7B-Instruct-64k-v0_1
81
+ - Finetuned from: [MediaTek-Research/Breeze-7B-Instruct-v0_1](https://huggingface.co/MediaTek-Research/Breeze-7B-Instruct-v0_1)
82
+ - Model type: Causal decoder-only transformer language model
83
+ - Language: English and Traditional Chinese (zh-tw)
84
+
85
+ ## Base Model Performance
86
+
87
+ **TMMLU+**, **DRCD**, and **Table** source from [MediaTek-Research/TCEval-v2](https://huggingface.co/datasets/MediaTek-Research/TCEval-v2).
88
+ [MediaTek-Research/TCEval-v2](https://huggingface.co/datasets/MediaTek-Research/TCEval-v2) derives from [TCEval-v1](https://github.com/mtkresearch/MR-Models/tree/main/TC-Eval)
89
+ and [ikala/tmmluplus](https://huggingface.co/datasets/ikala/tmmluplus). **MMLU** sources from [hails/mmlu_no_train](https://huggingface.co/datasets/hails/mmlu_no_train).
90
+ 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.
91
+
92
+
93
+ | Models | |↑ TMMLU+ (ACC) | DRCD (EM) | Table (ACC) | MMLU (ACC) |
94
+ |----------------------------------------------|--------|--------------|-------------|-------------|------------|
95
+ | | |TC, Knowledge |TC, Reasoning|TC, Reasoning|EN, Knowledge|
96
+ | | | 5 shot | 3 shot | 5 shot | 5 shot |
97
+ | [Yi-34B](https://huggingface.co/01-ai/Yi-34B)| 34B | 63.10 | 84.57 | 49.31 | 77.42 |
98
+ | [Qwen-14B](https://huggingface.co/01-ai/Qwen/Qwen-14B)| 14B | 51.30 | 16.95 * | 50.69 | 68.83 |
99
+ | [Yi-6B](https://huggingface.co/01-ai/Yi-6B) | 6B | 49.63 | 76.61 | 34.72 | 65.35 |
100
+ | [Qwen-7B](https://huggingface.co/01-ai/Qwen/Qwen-7B)| 7B | 42.84 | 0.0 * | 39.58 | 61.00 |
101
+ | [**Breeze-7B-Base-v0_1**](https://huggingface.co/MediaTek-Research/Breeze-7B-Base-v0_1) | 7B | 40.35 | 81.13 | 28.47 | 61.63 |
102
+ | [Mistral-7B-v0.1](https://huggingface.co/mistralai/Mistral-7B-v0.1)| 7B | 36.93 | 79.27 | 27.78 | 64.89 |
103
+
104
+
105
+ \* Few-shot learning cannot effectively guide the model to generate the proper answer.
106
+
107
+
108
+ ## Chat Model Performance
109
+
110
+ **TMMLU+**, **DRCD**, **Table**, and **MT-Bench-tw** source from [MediaTek-Research/TCEval-v2](https://huggingface.co/datasets/MediaTek-Research/TCEval-v2).
111
+ [MediaTek-Research/TCEval-v2](https://huggingface.co/datasets/MediaTek-Research/TCEval-v2) derives from [TCEval-v1](https://github.com/mtkresearch/MR-Models/tree/main/TC-Eval)
112
+ and [ikala/tmmluplus](https://huggingface.co/datasets/ikala/tmmluplus). **MMLU** sources from [hails/mmlu_no_train](https://huggingface.co/datasets/hails/mmlu_no_train).
113
+ **MT-Bench** source from [lmsys/mt_bench_human_judgments](https://huggingface.co/datasets/lmsys/mt_bench_human_judgments).
114
+ 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.
115
+ 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**.
116
+
117
+
118
+ | Models | |↑ MT-Bench-tw (Score)| TMMLU+ (ACC) | TMMLU+ (ACC) | DRCD (EM) | Table (ACC) | MT-Bench (Score) | MMLU (ACC) | MMLU (ACC) |
119
+ |---------------------------------------------------------------------------------------------------------|--------|--------------------|--------------|--------------|-------------|-------------|------------------|-------------|-------------|
120
+ | | |TC, Chat |TC, Knowledge |TC, Knowledge |TC, Reasoning|TC, Reasoning|EN, Chat |EN, Knowledge|EN, Knowledge|
121
+ | | |0 shot | 0 shot | 5 shot | 3 shot | 0 shot |0 shot | 0 shot | 5 shot |
122
+ | [gpt-3.5-turbo](https://openai.com) | |7.1 | 43.56 | | | 45.14 |7.9 | 67.09 | |
123
+ | [Yi-34B-Chat](https://huggingface.co/01-ai/Yi-34B-Chat) | 34B |6.9 | 54.87 | | | 36.81 |7.6 | 71.04 | |
124
+ | [Qwen-14B-Chat](https://huggingface.co/Qwen/Qwen-14B-Chat) | 14B |6.4 | 48.41 | | | 41.67 |7.2 | 64.91 | |
125
+ | [**Breeze-7B-Instruct-v0_1**](https://huggingface.co/MediaTek-Research/Breeze-7B-Instruct-v0_1) | 7B |5.7 | 41.61 | | | 45.83 |7.1 | 63.26 | |
126
+ | [**Breeze-7B-Instruct-64k-v0_1**](https://huggingface.co/MediaTek-Research/Breeze-7B-Instruct-64k-v0_1) | 7B |5.5 | 40.99 | | | 36.11 |7.1 | 63.68 | |
127
+ | [Qwen-7B-Chat](https://huggingface.co/Qwen/Qwen-7B-Chat) | 7B |5.4 | 40.02 | | | 33.33 |6.2 | 55.94 | |
128
+ | [Yi-6B-Chat](https://huggingface.co/01-ai/Yi-6B-Chat) | 6B |5.0 | 44.79 | | | 25.69 |6.0 | 59.45 | |
129
+ | [Taiwan-LLM-13B-v2.0-chat](https://huggingface.co/yentinglin/Taiwan-LLM-13B-v2.0-chat) | 13B |5.0 | 29.47 | | | 23.61 |-* | 50.50 | |
130
+ | [Taiwan-LLM-7B-v2.1-chat](https://huggingface.co/yentinglin/Taiwan-LLM-7B-v2.1-chat) | 7B |4.2 | 28.08 | | | 31.25 | -* | 42.72 | |
131
+
132
+ \* Taiwan-LLM models responds to multi-turn questions (English) in Traditional Chinese.
133
+
134
+
135
+ | Details on MT-Bench-tw (0 shot):<br/>Models | STEM |Extraction|Reasoning| Math | Coding | Roleplay| Writing |Humanities|↑ AVG |
136
+ |-----------------------------------------------------|---------|---------|---------|---------|---------|---------|---------|---------|---------|
137
+ | gpt-3.5-turbo | 7.8 | 6.1 | 5.1 | 6.4 | 6.2 | 8.7 | 7.4 | 9.3 | 7.1 |
138
+ | Yi-34B-Chat | 9.0 | 4.8 | 5.7 | 4.0 | 4.7 | 8.5 | 8.7 | 9.8 | 6.9 |
139
+ | Qwen-14B-Chat | 7.6 | 5.7 | 4.5 | 4.2 | 5.3 | 7.5 | 7.3 | 9.1 | 6.4 |
140
+ | **Breeze-7B-Instruct-v0_1** | 6.5 | 5.6 | 3.9 | 3.6 | 4.3 | 6.9 | 5.7 | 9.3 | 5.7 |
141
+ | **Breeze-7B-Instruct-64k-v0_1** | 6.1 | 5.3 | 3.7 | 2.9 | 4.2 | 7.0 | 6.7 | 8.3 | 5.5 |
142
+ | Qwen-7B-Chat | 6.6 | 4.5 | 4.8 | 2.9 | 3.6 | 6.2 | 6.8 | 8.2 | 5.4 |
143
+ | Yi-6B-Chat | 7.3 | 2.7 | 3.1 | 3.3 | 2.3 | 7.2 | 5.2 | 8.8 | 5.0 |
144
+ | Taiwan-LLM-13B-v2.0-chat | 6.1 | 3.4 | 4.1 | 2.3 | 3.1 | 7.4 | 6.6 | 6.8 | 5.0 |
145
+ | Taiwan-LLM-7B-v2.1-chat | 5.2 | 2.6 | 2.3 | 1.2 | 3.4 | 6.6 | 5.7 | 6.8 | 4.2 |
146
+
147
+
148
+ | Details on TMMLU+ (0 shot):<br/>Model | STEM | Social Science | Humanities | Other | ↑ AVG |
149
+ |-----------------------------------------------------|--------------|----------------|------------|------------|---------|
150
+ | Yi-34B-Chat | 47.65 | 64.25 | 52.73 | 54.91 | 54.87 |
151
+ | Qwen-14B-Chat | 43.83 | 55.00 | 48.55 | 46.22 | 48.41 |
152
+ | Yi-6B-Chat | 37.80 | 51.74 | 45.36 | 44.25 | 44.79 |
153
+ | gpt-3.5-turbo | 41.58 | 48.52 | 40.96 | 43.18 | 43.56 |
154
+ | **Breeze-7B-Instruct-v0_1** | 37.41 | 46.81 | 42.06 | 40.16 | 41.61 |
155
+ | **Breeze-7B-Instruct-64k-v0_1** | 37.88 | 46.35 | 40.31 | 39.40 | 40.99 |
156
+ | Qwen-7B-Chat | 35.44 | 46.22 | 38.35 | 40.06 | 40.02 |
157
+ | Taiwan-LLM-13B-v2.0-chat | 27.74 | 33.69 | 27.03 | 29.43 | 29.47 |
158
+ | Taiwan-LLM-7B-v2.1-chat | 25.58 | 31.76 | 27.36 | 27.61 | 28.08 |
159
+
160
+
161
+
162
+ ## Inference Performance
163
+ 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.
164
+ All inferences run on 2 RTX A6000 GPUs (using `vllm`, with a tensor-parallel size of 2).
165
+
166
+ | Models | ↓ Inference Time (sec)|Estimated Max Input Length (Char)|
167
+ |--------------------------------------------------------------------|-------------------|--------------------------|
168
+ | Yi-6B-Chat | 10.62 | 5.2k |
169
+ | **Breeze-7B-Instruct-v0_1** | 10.74 | 11.1k |
170
+ | **Breeze-7B-Instruct-64k-v0_1** | 10.74 | 88.8k |
171
+ | Qwen-7B-Chat | 10.86 | 9.8k |
172
+ | Qwen-14B-Chat | 18.89 | 9.8k |
173
+ | Mistral-7B-v0.1-Instruct | 20.48 | 5.1k |
174
+ | Taiwan-LLM-7B-v2.1-chat | 26.26 | 2.2k |
175
+ | Taiwan-LLM-13B-v2.0-chat | 36.80 | 2.2k |
176
+ | Yi-34B-Chat | 43.71 | 4.5k |
177
+
178
+ ## Long-context Performance
179
+
180
+ TBD
181
+
182
+ ## Use in Transformers
183
+
184
+ First install direct dependencies:
185
+ ```
186
+ pip install transformers torch accelerate
187
+ ```
188
+ If you want faster inference using flash-attention2, you need to install these dependencies:
189
+ ```bash
190
+ pip install packaging ninja
191
+ pip install flash-attn
192
+ ```
193
+ Then load the model in transformers:
194
+ ```python
195
+ from transformers import AutoModelForCausalLM, AutoTokenizer
196
+ import torch
197
+
198
+ model = AutoModelForCausalLM.from_pretrained(
199
+ "MediaTek-Research/Breeze-7B-Instruct-v0_1",
200
+ device_map="auto",
201
+ torch_dtype=torch.bfloat16,
202
+ attn_implementation="flash_attention_2" # optional
203
+ )
204
+ ```
205
+
206
+ The structure of the query is
207
+ ```txt
208
+ <s>SYS_PROMPT [INST] QUERY1 [/INST] RESPONSE1 [INST] QUERY2 [/INST]
209
+ ```
210
+ where `SYS_PROMPT`, `QUERY1`, `RESPONSE1`, and `QUERY2` can be provided by the user.
211
+
212
+ The suggested default `SYS_PROMPT` is
213
+ ```txt
214
+ You are a helpful AI assistant built by MediaTek Research. The user you are helping speaks Traditional Chinese and comes from Taiwan.
215
+ ```
216
+
217
+ We also integrate `chat_template` into [tokenizer_config.json](tokenizer_config.json), so you can `apply_chat_template` to get the prompt.
218
+
219
+ ```python
220
+ >>> from transformers import AutoTokenizer
221
+ >>> tokenizer = AutoTokenizer.from_pretrained("MediaTek-Research/Breeze-7B-Instruct-v0_1")
222
+ >>> chat = [
223
+ ... {"role": "user", "content": "你好，請問你可以完成什麼任務？"},
224
+ ... {"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。"},
225
+ ... {"role": "user", "content": "太棒了！"},
226
+ ... ]
227
+ >>> tokenizer.apply_chat_template(chat, tokenize=False)
228
+ "<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] "
229
+ # Tokenized results
230
+ # ['▁', '你好', '，', '請問', '你', '可以', '完成', '什麼', '任務', '？']
231
+ # ['▁', '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', '。']
232
+ # ['▁', '太', '棒', '了', '！']
233
+ ```
234
+
235
+ ## Citation
236
+
237
+ ```
238
+ @article{MediaTek-Research2024breeze7b,
239
+ title={Breeze-7B Technical Report},
240
+ author={Chan-Jan Hsu and Chang-Le Liu and Feng-Ting Liao and Po-Chun Hsu and Yi-Chang Chen and Da-Shan Shiu},
241
+ year={2024},
242
+ eprint={2403.02712},
243
+ archivePrefix={arXiv},
244
+ primaryClass={cs.CL}
245
+ }
246
+ ```
247
+