Taishi-N324 commited on
Commit
59eaa88
1 Parent(s): feb1a39

Upload README.md

Browse files
Files changed (1) hide show
  1. README.md +325 -0
README.md ADDED
@@ -0,0 +1,325 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ language:
3
+ - en
4
+ - ja
5
+ library_name: transformers
6
+ pipeline_tag: text-generation
7
+ license: llama2
8
+ model_type: llama
9
+ ---
10
+
11
+ # Swallow
12
+
13
+ Our Swallow model has undergone continuous pre-training from the [Llama 2 family](https://huggingface.co/meta-llama), primarily with the addition of Japanese language data. The tuned versions use supervised fine-tuning (SFT).
14
+ Links to other models can be found in the index.
15
+
16
+ # Model Release Updates
17
+
18
+ We are excited to share the release schedule for our latest models:
19
+ - **April 25, 2024**: Released version 1.0 of our enhanced instruction-tuned models: [Swallow-7b-instruct-v1.0](https://huggingface.co/tokyotech-llm/Swallow-7b-instruct-v1.0), [Swallow-13b-instruct-v1.0](https://huggingface.co/tokyotech-llm/Swallow-13b-instruct-v1.0), and [Swallow-70b-instruct-v1.0](https://huggingface.co/tokyotech-llm/Swallow-70b-instruct-v1.0).
20
+ - **March 2, 2024**: Released the [Swallow-7b-plus-hf](https://huggingface.co/tokyotech-llm/Swallow-7b-plus-hf), a model trained with approximately twice as many Japanese tokens as [Swallow-7b-hf](https://huggingface.co/tokyotech-llm/Swallow-7b-hf).
21
+ - **February 4, 2024**: Released the [Swallow-13b-NVE-hf](https://huggingface.co/tokyotech-llm/Swallow-13b-NVE-hf).
22
+ - **January 26, 2024**: Released the [Swallow-7b-NVE-hf](https://huggingface.co/tokyotech-llm/Swallow-7b-NVE-hf), [Swallow-7b-NVE-instruct-hf](https://huggingface.co/tokyotech-llm/Swallow-7b-NVE-instruct-hf), [Swallow-70b-NVE-hf](https://huggingface.co/tokyotech-llm/Swallow-70b-NVE-hf), and [Swallow-70b-NVE-instruct-hf](https://huggingface.co/tokyotech-llm/Swallow-70b-NVE-instruct-hf)
23
+ - **December 19, 2024**: Released the [Swallow-7b-hf](https://huggingface.co/tokyotech-llm/Swallow-7b-hf), [Swallow-7b-instruct-hf](https://huggingface.co/tokyotech-llm/Swallow-7b-instruct-hf), [Swallow-13b-hf](https://huggingface.co/tokyotech-llm/Swallow-13b-hf), [Swallow-13b-instruct-hf](https://huggingface.co/tokyotech-llm/Swallow-13b-instruct-hf), [Swallow-70b-hf](https://huggingface.co/tokyotech-llm/Swallow-70b-hf), and [Swallow-70b-instruct-hf](https://huggingface.co/tokyotech-llm/Swallow-70b-instruct-hf).
24
+ ## Swallow Model Index
25
+ |Model|Swallow-hf|Swallow-instruct-hf|Swallow-instruct-v1.0|
26
+ |---|---|---|---|
27
+ |7B| [Link](https://huggingface.co/tokyotech-llm/Swallow-7b-hf) | [Link](https://huggingface.co/tokyotech-llm/Swallow-7b-instruct-hf)|[Link](https://huggingface.co/tokyotech-llm/Swallow-7b-instruct-v1.0)|
28
+ |7B-Plus| [Link](https://huggingface.co/tokyotech-llm/Swallow-7b-plus-hf) | N/A | N/A |
29
+ |13B| [Link](https://huggingface.co/tokyotech-llm/Swallow-13b-hf) | [Link](https://huggingface.co/tokyotech-llm/Swallow-13b-instruct-hf)| [Link](https://huggingface.co/tokyotech-llm/Swallow-13b-instruct-v1.0)|
30
+ |70B| [Link](https://huggingface.co/tokyotech-llm/Swallow-70b-hf) | [Link](https://huggingface.co/tokyotech-llm/Swallow-70b-instruct-hf)| [Link](https://huggingface.co/tokyotech-llm/Swallow-70b-instruct-v1.0)|
31
+
32
+ ## Swallow Model Index NVE (No Vocabulary Expansion)
33
+ |Model|Swallow-NVE-hf|Swallow-NVE-instruct-hf|
34
+ |---|---|---|
35
+ |7B| [Link](https://huggingface.co/tokyotech-llm/Swallow-7b-NVE-hf) | [Link](https://huggingface.co/tokyotech-llm/Swallow-7b-NVE-instruct-hf)|
36
+ |13B| [Link](https://huggingface.co/tokyotech-llm/Swallow-13b-NVE-hf) | N/A |
37
+ |70B| [Link](https://huggingface.co/tokyotech-llm/Swallow-70b-NVE-hf) | [Link](https://huggingface.co/tokyotech-llm/Swallow-70b-NVE-instruct-hf)|
38
+
39
+ ![logo](./logo.png)
40
+
41
+ This repository provides large language models developed by [TokyoTech-LLM](https://tokyotech-llm.github.io/).
42
+ Read our [blog post](https://zenn.dev/tokyotech_lm/articles/d6cb3a8fdfc907) or our [paper](https://www.anlp.jp/proceedings/annual_meeting/2024/pdf_dir/A8-5.pdf)
43
+
44
+ ## Model Details
45
+
46
+ * **Model type**: Please refer to LLaMA-2 technical report for details on the model architecture.
47
+ * **Language(s)**: Japanese English
48
+ * **Library**: [Megatron-LM](https://github.com/rioyokotalab/Megatron-Llama2)
49
+ * **Tokenizer**: This model employs a tokenizer that features a broadened vocabulary based on Japanese data. This allows for a more efficient representation of text using fewer tokens, leading to a notably faster inference process.
50
+ * **Contact**: swallow[at]nlp.c.titech.ac.jp
51
+
52
+ ## Instruct Model Performance
53
+
54
+ ### MT-Bench JA
55
+
56
+ TODO
57
+
58
+ ## Base Model Performance
59
+
60
+ ### Japanese tasks
61
+
62
+ |Model|Size|JCommonsenseQA|JEMHopQA|NIILC|JSQuAD|XL-Sum|MGSM|WMT20-en-ja|WMT20-ja-en|
63
+ |---|---|---|---|---|---|---|---|---|---|
64
+ | | |4-shot|4-shot|4-shot|4-shot|1-shot|4-shot|4-shot|4-shot|
65
+ | Llama 2 | 7B | 0.3852 | 0.4240 | 0.3410 | 0.7917 | 0.1905 | 0.0760 | 0.1783 | 0.1738 |
66
+ | Swallow | 7B | 0.4808 | 0.5078 | 0.5968 | 0.8573 | 0.1830 | 0.1240 | 0.2510 | 0.1511 |
67
+ | Swallow-Plus | 7B | 0.5478 | 0.5493 | 0.6030 | 0.8544 | 0.1806 | 0.1360 | 0.2568 | 0.1441 |
68
+ | Swallow-NVE | 7B | 0.5433 | 0.5425 | 0.5729 | 0.8684 | 0.2117 | 0.1200 | 0.2405 | 0.1512 |
69
+ | Llama 2 | 13B | 0.6997 | 0.4415 | 0.4170 | 0.8533 | 0.2139 | 0.1320 | 0.2146 | 0.1982 |
70
+ | Swallow | 13B | 0.7837 | 0.5063 | 0.6398 | 0.9005 | 0.2168 | 0.2040 | 0.2720 | 0.1771 |
71
+ | Swallow-NVE | 13B | 0.7712 | 0.5438 | 0.6351 | 0.9030 | 0.2294 | 0.2120 | 0.2735 | 0.1817 |
72
+ | Llama 2 | 70B | 0.8686 | 0.4656 | 0.5256 | 0.9080 | 0.2361 | 0.3560 | 0.2643 | **0.2398** |
73
+ | Swallow | 70B | 0.9348 | **0.6290** | 0.6960 | 0.9176 | 0.2266 | **0.4840** | **0.3043** | 0.2298 |
74
+ | Swallow-NVE | 70B | **0.9410** | 0.5759 | **0.7024** | **0.9254** | **0.2758** | 0.4720 | 0.3042 | 0.2322 |
75
+ ### English tasks
76
+
77
+ |Model|Size|OpenBookQA|TriviaQA|HellaSwag|SQuAD2.0|XWINO|GSM8K|
78
+ |---|---|---|---|---|---|---|---|
79
+ | | |8-shot|8-shot|8-shot|8-shot|8-shot|8-shot|
80
+ | Llama 2 | 7B | 0.3580 | 0.6265 | 0.5860 | 0.3207 | 0.9049 | 0.1410 |
81
+ | Swallow | 7B | 0.3180 | 0.4836 | 0.5308 | 0.3125 | 0.8817 | 0.1130 |
82
+ | Swallow-Plus | 7B | 0.3280 | 0.4558 | 0.5259 | 0.3134 | 0.8929 | 0.1061 |
83
+ | Swallow-NVE | 7B | 0.3180 | 0.5079 | 0.5329 | 0.2919 | 0.8817 | 0.0986 |
84
+ | Llama 2 | 13B | 0.3760 | 0.7255 | 0.6148 | 0.3681 | 0.9140 | 0.2403 |
85
+ | Swallow | 13B | 0.3500 | 0.5852 | 0.5660 | 0.3406 | 0.9075 | 0.2039 |
86
+ | Swallow-NVE | 13B | 0.3460 | 0.6025 | 0.5700 | 0.3478 | 0.9006 | 0.1751 |
87
+ | Llama 2 | 70B | **0.4280** | **0.8239** | **0.6742** | **0.3770** | **0.9290** | **0.5284** |
88
+ | Swallow | 70B | 0.4220 | 0.7756 | 0.6458 | 0.3745 | 0.9204 | 0.4867 |
89
+ | Swallow-NVE | 70B | 0.4240 | 0.7817 | 0.6439 | 0.3451 | 0.9256 | 0.4943 |
90
+
91
+ ## Evaluation Benchmarks
92
+
93
+ ### Japanese evaluation benchmarks
94
+
95
+ We used llm-jp-eval(v1.0.0) and JP Language Model Evaluation Harness(commit #9b42d41). The details are as follows:
96
+
97
+ - Multiple-choice question answering (JCommonsenseQA [Kurihara+, 2022])
98
+ - Open-ended question answering (JEMHopQA [Ishii+, 2023])
99
+ - Open-ended question answering (NIILC [Sekine, 2003])
100
+ - Machine reading comprehension (JSQuAD [Kurihara+, 2022])
101
+ - Automatic summarization (XL-Sum [Hasan+, 2021])
102
+ - Machine translation (WMT2020 ja-en [Barrault+, 2020])
103
+ - Machine translation (WMT2020 en-ja [Barrault+, 2020])
104
+ - Mathematical reasoning (MGSM [Shi+, 2023])
105
+
106
+ ### English evaluation benchmarks
107
+
108
+ We used the Language Model Evaluation Harness(v.0.3.0). The details are as follows:
109
+
110
+ - Multiple-choice question answering (OpenBookQA [Mihaylov+, 2018])
111
+ - Open-ended question answering (TriviaQA [Joshi+, 2017])
112
+ - Machine reading comprehension (SQuAD 2.0 [Rajpurkar+, 2018])
113
+ - Commonsense reasoning (XWINO [Tikhonov & Ryabinin, 2021])
114
+ - Natural language inference (HellaSwag [Zellers+, 2019])
115
+ - Mathematical reasoning (GSM8k [Cobbe+, 2021])
116
+
117
+
118
+ ## Usage
119
+
120
+ First install additional dependencies in [requirements.txt](./requirements.txt):
121
+
122
+ ```sh
123
+ pip install -r requirements.txt
124
+ ```
125
+
126
+ ### Instruction format Ver1.0
127
+ This format must be adhered to strictly, as deviations may result in less optimal outputs from the model.
128
+
129
+ The template used to construct a prompt for the Instruct model is specified as follows:
130
+
131
+ ```
132
+ <s>[INST] <<SYS>>\n{Instruction}\n<</SYS>>\n\n{USER_MESSAGE_1} [INST] {BOT_MESSAGE_1} </s>[INST] {USER_MESSAGE_2}[/INST]
133
+ ```
134
+
135
+ Please be aware that <s> and </s> are special tokens used for the beginning of string (BOS) and end of string (EOS), respectively, while [INST] and [/INST] are considered regular strings.
136
+
137
+ ### Use the instruct model Ver1.0
138
+
139
+ ```python
140
+ import torch
141
+ from transformers import AutoTokenizer, AutoModelForCausalLM
142
+
143
+ model_name = "tokyotech-llm/Swallow-70b-instruct-v1.0"
144
+ model = AutoModelForCausalLM.from_pretrained(model_name, torch_dtype=torch.bfloat16, device_map="auto")
145
+ tokenizer = AutoTokenizer.from_pretrained(model_name)
146
+
147
+ device = "cuda"
148
+
149
+ messages = [
150
+ {"role": "system", "content": "あなたは誠実で優秀な日本人のアシスタントです。"},
151
+ {"role": "user", "content": "東京工業大学の主なキャンパスについて教えてください"}
152
+ ]
153
+
154
+ encodeds = tokenizer.apply_chat_template(messages, return_tensors="pt")
155
+
156
+ model_inputs = encodeds.to(device)
157
+ model.to(device)
158
+
159
+ generated_ids = model.generate(model_inputs, max_new_tokens=128, do_sample=True)
160
+ decoded = tokenizer.batch_decode(generated_ids)
161
+ print(decoded[0])
162
+ ```
163
+
164
+ ### Use the instruct model
165
+
166
+ **Note:** Please be aware that the inference example is based on a model version older than 1.0.
167
+
168
+
169
+ ```python
170
+ import torch
171
+ from transformers import AutoTokenizer, AutoModelForCausalLM
172
+
173
+ model_name = "tokyotech-llm/Swallow-7b-instruct-hf"
174
+
175
+ tokenizer = AutoTokenizer.from_pretrained(model_name)
176
+ model = AutoModelForCausalLM.from_pretrained(model_name, torch_dtype=torch.bfloat16, low_cpu_mem_usage=True, device_map="auto")
177
+
178
+
179
+ PROMPT_DICT = {
180
+ "prompt_input": (
181
+ "以下に、あるタスクを説明する指示があり、それに付随する入力が更なる文脈を提供しています。"
182
+ "リクエストを適切に完了するための回答を記述してください。\n\n"
183
+ "### 指示:\n{instruction}\n\n### 入力:\n{input}\n\n### 応答:"
184
+
185
+ ),
186
+ "prompt_no_input": (
187
+ "以下に、あるタスクを説明する指示があります。"
188
+ "リクエストを適切に完了するための回答を記述してください。\n\n"
189
+ "### 指示:\n{instruction}\n\n### 応答:"
190
+ ),
191
+ }
192
+
193
+ def create_prompt(instruction, input=None):
194
+ """
195
+ Generates a prompt based on the given instruction and an optional input.
196
+ If input is provided, it uses the 'prompt_input' template from PROMPT_DICT.
197
+ If no input is provided, it uses the 'prompt_no_input' template.
198
+
199
+ Args:
200
+ instruction (str): The instruction describing the task.
201
+ input (str, optional): Additional input providing context for the task. Default is None.
202
+
203
+ Returns:
204
+ str: The generated prompt.
205
+ """
206
+ if input:
207
+ # Use the 'prompt_input' template when additional input is provided
208
+ return PROMPT_DICT["prompt_input"].format(instruction=instruction, input=input)
209
+ else:
210
+ # Use the 'prompt_no_input' template when no additional input is provided
211
+ return PROMPT_DICT["prompt_no_input"].format(instruction=instruction)
212
+
213
+ # Example usage
214
+ instruction_example = "以下のトピックに関する詳細な情報を提供してください。"
215
+ input_example = "東京工業大学の主なキャンパスについて教えてください"
216
+ prompt = create_prompt(instruction_example, input_example)
217
+
218
+ input_ids = tokenizer.encode(
219
+ prompt,
220
+ add_special_tokens=False,
221
+ return_tensors="pt"
222
+ )
223
+
224
+ tokens = model.generate(
225
+ input_ids.to(device=model.device),
226
+ max_new_tokens=128,
227
+ temperature=0.99,
228
+ top_p=0.95,
229
+ do_sample=True,
230
+ )
231
+
232
+ out = tokenizer.decode(tokens[0], skip_special_tokens=True)
233
+ print(out)
234
+
235
+ ```
236
+
237
+ ### Use the base model
238
+
239
+ ```python
240
+ import torch
241
+ from transformers import AutoTokenizer, AutoModelForCausalLM
242
+
243
+ model_name = "tokyotech-llm/Swallow-7b-hf"
244
+
245
+ tokenizer = AutoTokenizer.from_pretrained(model_name)
246
+ model = AutoModelForCausalLM.from_pretrained(model_name, torch_dtype=torch.bfloat16, device_map="auto")
247
+
248
+ prompt = "東京工業大学の主なキャンパスは、"
249
+ input_ids = tokenizer.encode(
250
+ prompt,
251
+ add_special_tokens=False,
252
+ return_tensors="pt"
253
+ )
254
+ tokens = model.generate(
255
+ input_ids.to(device=model.device),
256
+ max_new_tokens=128,
257
+ temperature=0.99,
258
+ top_p=0.95,
259
+ do_sample=True,
260
+ )
261
+
262
+ out = tokenizer.decode(tokens[0], skip_special_tokens=True)
263
+ print(out)
264
+ ```
265
+
266
+ ## Training Datasets
267
+
268
+ ### Continual Pre-Training
269
+ The following datasets were used for continual pre-training.
270
+
271
+ - [Japanese Wikipedia](https://dumps.wikimedia.org/other/cirrussearch)
272
+ - [RefinedWeb](https://huggingface.co/datasets/tiiuae/falcon-refinedweb)
273
+ - Swallow Corpus
274
+ - [The Pile](https://huggingface.co/datasets/EleutherAI/pile)
275
+
276
+
277
+ ### Instruction Tuning
278
+
279
+ #### Ver1.0
280
+
281
+ The following datasets were used for the instruction tuning.
282
+
283
+ - [OpenAssistant Conversations Dataset EN top-1 thread](https://huggingface.co/datasets/OpenAssistant/oasst2)
284
+ - [OpenAssistant Conversations Dataset](https://huggingface.co/datasets/llm-jp/oasst1-21k-ja) was used, where human utterances are included but the responses are not used. Instead, the responses were generated using the [Mixtral-8x7B-Instruct-v0.1](https://huggingface.co/datasets/llm-jp/oasst1-21k-jahttps://huggingface.co/mistralai/Mixtral-8x7B-Instruct-v0.1) model.
285
+
286
+ #### Old
287
+
288
+ The following datasets were used for the instruction tuning.
289
+
290
+ - [Anthropic HH-RLHF](https://huggingface.co/datasets/kunishou/hh-rlhf-49k-ja)
291
+ - [Databricks Dolly 15-k](https://huggingface.co/datasets/kunishou/databricks-dolly-15k-ja)
292
+ - [OpenAssistant Conversations Dataset](https://huggingface.co/datasets/kunishou/oasst1-89k-ja)
293
+
294
+ ## Risks and Limitations
295
+
296
+ The models released here are still in the early stages of our research and development and have not been tuned to ensure outputs align with human intent and safety considerations.
297
+
298
+ ## Acknowledgements
299
+
300
+ We thank Meta Research for releasing Llama 2 under an open license for others to build on.
301
+
302
+ Our project is supported by the [ABCI Large-scale Language Model Building Support Program](https://abci.ai/en/link/llm_support_program.html) of the National Institute of Advanced Industrial Science and Technology.
303
+
304
+ ## License
305
+
306
+ Llama 2 is licensed under the LLAMA 2 Community License, Copyright © Meta Platforms, Inc. All Rights Reserved.
307
+
308
+ ## Authors
309
+
310
+ Here are the team members:
311
+ - From [Okazaki Laboratory](https://www.nlp.c.titech.ac.jp/index.en.html), the following members:
312
+ - [Naoaki Okazaki](https://www.chokkan.org/index.ja.html)
313
+ - [Sakae Mizuki](https://s-mizuki-nlp.github.io/)
314
+ - [Hiroki Iida](https://meshidenn.github.io/)
315
+ - [Mengsay Loem](https://loem-ms.github.io/)
316
+ - [Shota Hirai](https://huggingface.co/Kotemo428)
317
+ - [Kakeru Hattori](https://aya-se.vercel.app/)
318
+ - [Masanari Ohi](https://twitter.com/stjohn2007)
319
+ - From [YOKOTA Laboratory](https://www.rio.gsic.titech.ac.jp/en/index.html), the following members:
320
+ - [Rio Yokota](https://twitter.com/rioyokota)
321
+ - [Kazuki Fujii](https://twitter.com/okoge_kaz)
322
+ - [Taishi Nakamura](https://twitter.com/Setuna7777_2)
323
+ - [Takumi Okamoto](https://www.linkedin.com/in/takumi-okamoto)
324
+ - [Ishida Shigeki](https://www.wantedly.com/id/reborn27)
325
+