sho-takase
commited on
Commit
•
bf3ad25
1
Parent(s):
fdb7188
Update README.md
Browse files
README.md
CHANGED
@@ -1,71 +1,71 @@
|
|
1 |
-
---
|
2 |
-
license: mit
|
3 |
-
language:
|
4 |
-
- ja
|
5 |
-
- en
|
6 |
-
---
|
7 |
-
|
8 |
-
# Sarashina2-7B
|
9 |
-
|
10 |
-
This repository provides large language models trained by [SB Intuitions](https://www.sbintuitions.co.jp/).
|
11 |
-
|
12 |
-
|
13 |
-
## How to use
|
14 |
-
|
15 |
-
```
|
16 |
-
import torch
|
17 |
-
from transformers import AutoModelForCausalLM, AutoTokenizer, pipeline, set_seed
|
18 |
-
|
19 |
-
model = AutoModelForCausalLM.from_pretrained("sbintuitions/sarashina2-7b", torch_dtype=torch.bfloat16)
|
20 |
-
tokenizer = AutoTokenizer.from_pretrained("sbintuitions/sarashina2-7b", use_fast=False)
|
21 |
-
generator = pipeline("text-generation", model=model, tokenizer=tokenizer
|
22 |
-
set_seed(123)
|
23 |
-
|
24 |
-
text = generator(
|
25 |
-
"おはようございます、今日の天気は",
|
26 |
-
max_length=30,
|
27 |
-
do_sample=True,
|
28 |
-
pad_token_id=tokenizer.pad_token_id,
|
29 |
-
num_return_sequences=3,
|
30 |
-
)
|
31 |
-
|
32 |
-
for t in text:
|
33 |
-
print(t)
|
34 |
-
|
35 |
-
# These examples are generated by sarashina2-7b parameters model
|
36 |
-
# {'generated_text': 'おはようございます、今日の天気は晴れです。ちょっと風が強い。\n昨日は、久しぶりにゆっくりとしていました。\n2週間位間があいてしまったかも、でもその間に'}
|
37 |
-
# {'generated_text': 'おはようございます、今日の天気は曇。朝は曇っていてどんよりしていましたね。昼からは晴れそうですが。気温は徐々に上昇しています。昨日は春らしい陽気でした。'}
|
38 |
-
# {'generated_text': 'おはようございます、今日の天気はくもり、少し寒気がします。
|
39 |
-
```
|
40 |
-
|
41 |
-
## Configuration
|
42 |
-
|
43 |
-
| Parameters | Vocab size | Trainning tokens | Architecture | Position type | Layers | Hidden dim | Attention heads |
|
44 |
-
| :-----: | :-----------: | :-------------: | :------------ | :-----------: | :----: | :--------: | :-------------: |
|
45 |
-
| [7B](https://huggingface.co/sbintuitions/sarashina2-7b) | 102400 | 2.1T | Llama2 | RoPE | 32 | 4096 | 32 |
|
46 |
-
| [13B](https://huggingface.co/sbintuitions/sarashina2-13b) | 102400 | 2.1T | Llama2 | RoPE | 40 | 5120 | 40 |
|
47 |
-
| 70B (TBA)| | | | | | |
|
48 |
-
|
49 |
-
## Training Corpus
|
50 |
-
|
51 |
-
For our Japanese training data, we used a Japanese portion of the [Common Crawl corpus](https://commoncrawl.org/), which is the largest Web corpus, as our training dataset.
|
52 |
-
To clean the training corpus, we used [CCNet](https://github.com/facebookresearch/cc_net) and [HojiChar](https://github.com/HojiChar/HojiChar).
|
53 |
-
After cleaning, our Japanese training data contains about 1T tokens.
|
54 |
-
|
55 |
-
For our English training data, we extracted English documents from [SlimPajama](https://huggingface.co/datasets/cerebras/SlimPajama-627B) but we removed books3 corpus due to copyright infringement.
|
56 |
-
|
57 |
-
## Tokenization
|
58 |
-
|
59 |
-
We use a [sentencepiece](https://github.com/google/sentencepiece) tokenizer with a unigram language model and byte-fallback.
|
60 |
-
We do not apply pre-tokenization with Japanese tokenizer.
|
61 |
-
Thus, a user may directly feed raw sentences into the tokenizer.
|
62 |
-
|
63 |
-
|
64 |
-
## Ethical Considerations and Limitations
|
65 |
-
Sarashina2 has not been tuned to follow an instruction yet.
|
66 |
-
Therefore, sarashina2 might generate some meaningless sequences, some inaccurate instances or biased/objectionable outputs.
|
67 |
-
Before using sarashina2, we would like developers to tune models based on human preferences and safety considerations.
|
68 |
-
|
69 |
-
## License
|
70 |
-
|
71 |
[MIT License](https://huggingface.co/sbintuitions/sarashina2-7b/blob/main/LICENSE)
|
|
|
1 |
+
---
|
2 |
+
license: mit
|
3 |
+
language:
|
4 |
+
- ja
|
5 |
+
- en
|
6 |
+
---
|
7 |
+
|
8 |
+
# Sarashina2-7B
|
9 |
+
|
10 |
+
This repository provides large language models trained by [SB Intuitions](https://www.sbintuitions.co.jp/).
|
11 |
+
|
12 |
+
|
13 |
+
## How to use
|
14 |
+
|
15 |
+
```
|
16 |
+
import torch
|
17 |
+
from transformers import AutoModelForCausalLM, AutoTokenizer, pipeline, set_seed
|
18 |
+
|
19 |
+
model = AutoModelForCausalLM.from_pretrained("sbintuitions/sarashina2-7b", torch_dtype=torch.bfloat16, device_map="auto")
|
20 |
+
tokenizer = AutoTokenizer.from_pretrained("sbintuitions/sarashina2-7b", use_fast=False)
|
21 |
+
generator = pipeline("text-generation", model=model, tokenizer=tokenizer)
|
22 |
+
set_seed(123)
|
23 |
+
|
24 |
+
text = generator(
|
25 |
+
"おはようございます、今日の天気は",
|
26 |
+
max_length=30,
|
27 |
+
do_sample=True,
|
28 |
+
pad_token_id=tokenizer.pad_token_id,
|
29 |
+
num_return_sequences=3,
|
30 |
+
)
|
31 |
+
|
32 |
+
for t in text:
|
33 |
+
print(t)
|
34 |
+
|
35 |
+
# These examples are generated by sarashina2-7b parameters model
|
36 |
+
# {'generated_text': 'おはようございます、今日の天気は晴れです。ちょっと風が強い。\n昨日は、久しぶりにゆっくりとしていました。\n2週間位間があいてしまったかも、でもその間に'}
|
37 |
+
# {'generated_text': 'おはようございます、今日の天気は曇。朝は曇っていてどんよりしていましたね。昼からは晴れそうですが。気温は徐々に上昇しています。昨日は春らしい陽気でした。'}
|
38 |
+
# {'generated_text': 'おはようございます、今日の天気はくもり、少し寒気がします。 この土日に、家族で一泊��日で旅行に行ってきました。といっても、100キロ'}
|
39 |
+
```
|
40 |
+
|
41 |
+
## Configuration
|
42 |
+
|
43 |
+
| Parameters | Vocab size | Trainning tokens | Architecture | Position type | Layers | Hidden dim | Attention heads |
|
44 |
+
| :-----: | :-----------: | :-------------: | :------------ | :-----------: | :----: | :--------: | :-------------: |
|
45 |
+
| [7B](https://huggingface.co/sbintuitions/sarashina2-7b) | 102400 | 2.1T | Llama2 | RoPE | 32 | 4096 | 32 |
|
46 |
+
| [13B](https://huggingface.co/sbintuitions/sarashina2-13b) | 102400 | 2.1T | Llama2 | RoPE | 40 | 5120 | 40 |
|
47 |
+
| 70B (TBA)| | | | | | |
|
48 |
+
|
49 |
+
## Training Corpus
|
50 |
+
|
51 |
+
For our Japanese training data, we used a Japanese portion of the [Common Crawl corpus](https://commoncrawl.org/), which is the largest Web corpus, as our training dataset.
|
52 |
+
To clean the training corpus, we used [CCNet](https://github.com/facebookresearch/cc_net) and [HojiChar](https://github.com/HojiChar/HojiChar).
|
53 |
+
After cleaning, our Japanese training data contains about 1T tokens.
|
54 |
+
|
55 |
+
For our English training data, we extracted English documents from [SlimPajama](https://huggingface.co/datasets/cerebras/SlimPajama-627B) but we removed books3 corpus due to copyright infringement.
|
56 |
+
|
57 |
+
## Tokenization
|
58 |
+
|
59 |
+
We use a [sentencepiece](https://github.com/google/sentencepiece) tokenizer with a unigram language model and byte-fallback.
|
60 |
+
We do not apply pre-tokenization with Japanese tokenizer.
|
61 |
+
Thus, a user may directly feed raw sentences into the tokenizer.
|
62 |
+
|
63 |
+
|
64 |
+
## Ethical Considerations and Limitations
|
65 |
+
Sarashina2 has not been tuned to follow an instruction yet.
|
66 |
+
Therefore, sarashina2 might generate some meaningless sequences, some inaccurate instances or biased/objectionable outputs.
|
67 |
+
Before using sarashina2, we would like developers to tune models based on human preferences and safety considerations.
|
68 |
+
|
69 |
+
## License
|
70 |
+
|
71 |
[MIT License](https://huggingface.co/sbintuitions/sarashina2-7b/blob/main/LICENSE)
|