sho-takase
commited on
Commit
•
24e42c1
1
Parent(s):
eabe8fa
Add readme
Browse files
README.md
ADDED
@@ -0,0 +1,38 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
---
|
2 |
+
language:
|
3 |
+
- ja
|
4 |
+
- en
|
5 |
+
---
|
6 |
+
|
7 |
+
# Sarashina2-8x70B
|
8 |
+
|
9 |
+
This repository provides large language models trained by [SB Intuitions](https://www.sbintuitions.co.jp/).
|
10 |
+
|
11 |
+
|
12 |
+
## Required Hardware
|
13 |
+
BF16 Inference:
|
14 |
+
- 16x H100
|
15 |
+
- 16x A100 80GB
|
16 |
+
|
17 |
+
|
18 |
+
## Model Description
|
19 |
+
|
20 |
+
We constructed this Sarashina2-8x70B model, which consists of over 450 billion parameters, by applying the [sparse upcycling technique](https://arxiv.org/abs/2212.05055) to our [Sarashina2-70B](https://huggingface.co/sbintuitions/sarashina2-70b) model to efficiently build the Mixture-of-Experts model.
|
21 |
+
We trained the Sarashina2-8x70B model using a mix of Japanese and English corpora from web data.
|
22 |
+
|
23 |
+
|
24 |
+
## Tokenization
|
25 |
+
|
26 |
+
We use a [sentencepiece](https://github.com/google/sentencepiece) tokenizer with a unigram language model and byte-fallback.
|
27 |
+
We do not apply pre-tokenization with Japanese tokenizer.
|
28 |
+
Thus, a user may directly feed raw sentences into the tokenizer.
|
29 |
+
|
30 |
+
|
31 |
+
## Ethical Considerations and Limitations
|
32 |
+
Sarashina2 has not been tuned to follow an instruction yet.
|
33 |
+
Therefore, sarashina2 might generate some meaningless sequences, some inaccurate instances or biased/objectionable outputs.
|
34 |
+
Before using sarashina2, we would like developers to tune models based on human preferences and safety considerations.
|
35 |
+
|
36 |
+
## License
|
37 |
+
|
38 |
+
[Sarashina Model NonCommercial License Agreement](https://huggingface.co/sbintuitions/sarashina2-8x70B/blob/main/Sarashina%20Model%20NonCommercial%20License%20Agreement)
|