Create README.md
Browse files
README.md
ADDED
@@ -0,0 +1,84 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
---
|
2 |
+
language:
|
3 |
+
- jp
|
4 |
+
tags:
|
5 |
+
- pytorch
|
6 |
+
- causal-lm
|
7 |
+
license: apache-2.0
|
8 |
+
|
9 |
+
---
|
10 |
+
|
11 |
+
# Genji-JP 6B
|
12 |
+
|
13 |
+
Please check our blog post for more details, samples, evaluations and more:
|
14 |
+
[Blogpost](https://colab.research.google.com/drive/1PnWpx02IEUkY8jhLKd_NewUGEXahAska?usp=sharing)
|
15 |
+
|
16 |
+
## Model Description
|
17 |
+
|
18 |
+
Genji-JP 6B is a model finetuned on our Japanese storytelling dataset based off EleutherAI's GPT-J 6B model. This particular model is trained on Japanese web novels.
|
19 |
+
|
20 |
+
| Hyperparameter | Value |
|
21 |
+
|-------------------|--------|
|
22 |
+
| n_parameters | 6,053,381,344 |
|
23 |
+
| n_layers | 28* |
|
24 |
+
| d_model | 4,096 |
|
25 |
+
| d_ff | 16,384 |
|
26 |
+
| n_heads | 16 |
|
27 |
+
| d_head | 256 |
|
28 |
+
| n_ctx | 2,048 |
|
29 |
+
| n_vocab | 50,400 (same tokenizer as GPT-2/3) |
|
30 |
+
| position encoding | [Rotary position encodings (RoPE)](https://arxiv.org/abs/2104.09864) |
|
31 |
+
| RoPE dimensions | [64](https://github.com/kingoflolz/mesh-transformer-jax/blob/f2aa66e0925de6593dcbb70e72399b97b4130482/mesh_transformer/layers.py#L223) |
|
32 |
+
|
33 |
+
`*` each layer consists of one feedforward block and one self attention block
|
34 |
+
|
35 |
+
The model consists of 28 layers with a model dimension of 4096, and a feedforward dimension of 16384. The model
|
36 |
+
dimension is split into 16 heads, each with a dimension of 256. Rotary position encodings (RoPE) was applied to 64
|
37 |
+
dimensions of each head. The model is trained with a tokenization vocabulary of 50257, using the same set of BPEs as
|
38 |
+
GPT-2/GPT-3.
|
39 |
+
|
40 |
+
## Training data
|
41 |
+
|
42 |
+
GPT-J 6B was pretrained on the [Pile](pile.eleuther.ai), a large scale curated dataset created by EleutherAI for the purpose of training this model. After the pre-training, it's finetuned on the python code that was taken from the Pile.
|
43 |
+
|
44 |
+
### How to use
|
45 |
+
|
46 |
+
```from transformers import AutoTokenizer, AutoModelForCausalLM
|
47 |
+
import torch
|
48 |
+
|
49 |
+
tokenizer = AutoTokenizer.from_pretrained("EleutherAI/gpt-j-6B")
|
50 |
+
model = AutoModelForCausalLM.from_pretrained("NovelAI/genji-jp", torch_dtype=torch.float16, low_cpu_mem_usage=True).eval().cuda()
|
51 |
+
text = '''ใใใใ๏ผใใชใใฏ็ฐไธ็ใซ่ปข็ใใฆใใพใใพใใใๅ่
ใจใชใฃใฆใไปฒ้ใไฝใใ็ฐไธ็ใๅ้บใใใ๏ผ
|
52 |
+
***
|
53 |
+
่ปข็ใใใจใใใ่ฝๅใๆใซๅ
ฅใใฆใใใใใใฏใ'''
|
54 |
+
|
55 |
+
tokens = tokenizer(text, return_tensors="pt").input_ids
|
56 |
+
generated_tokens = model.generate(tokens.long().cuda(), use_cache=True, do_sample=True, temperature=1, top_p=0.9, repetition_penalty=1.125, min_length=1, max_length=len(tokens[0]) + 400, pad_token_id=tokenizer.eos_token_id)
|
57 |
+
last_tokens = generated_tokens[0]
|
58 |
+
generated_text = tokenizer.decode(last_tokens).replace("๏ฟฝ", "")
|
59 |
+
print("Generation:\n" + generated_text)
|
60 |
+
```
|
61 |
+
When run, this code generates:
|
62 |
+
```
|
63 |
+
Generation:
|
64 |
+
ใใใใ๏ผใใชใใฏ็ฐไธ็ใซ่ปข็ใใฆใใพใใพใใใๅ่
ใจใชใฃใฆใไปฒ้ใไฝใใ็ฐไธ็ใๅ้บใใใ๏ผ
|
65 |
+
***
|
66 |
+
่ปข็ใใใจใใใ่ฝๅใๆใซๅ
ฅใใฆใใใใใใฏใใไบ็ฅใใ ใ้ๅปใใๆชๆฅใฎใใจใใ่ชฐใ็ฅใใชใๅบๆฅไบใๅซใใฆ่ฆ้ใใใจใๅบๆฅใใ
|
67 |
+
ๆช้ญใฎๆฌ ็ใจๅผใฐใใๅฐใใช็ตๆถใๅใ่พผใใงใไฝฟๅฝนใใใใจใๅบๆฅใใไบบใๆนใใคใใๅ ่ฝใใใใไฝใใใไฟบใฏ็ทใชใใฆๅฑ
ใชใใฃใใใๅฅณใซ่ๅณใใชใใโฆโฆใใใชใฏใบใฎ็ๆฃใๆ
ใไธใใๅฅดใๅคใใชใใจๆใใจใใกใใฃใจ่ฆใใใ
|
68 |
+
ใ ใใไธ้จใฎไบบ้ใซใฏๅๅ่
ใๅพใใใจใๅบๆฅใใ็ฎ็ซใใชใ่กใซใใๅฏบใฎไธญใงใๅธธใซๅฎถใซๅผใใใใฃใฆใใ่ไบบใใใใชใคใใฎ้ญใใณใณใใญใผใซใใใใจใๅบๆฅใใฎใ ใไพฟๅฉใช่ฝๅใ ใใใใใ่ฃๅใ่
ใฏๅคงๅขใใใๆฐใๆใใฐใ็ใใใ ใใๆณจๆใๅฟ
่ฆใ ใ
|
69 |
+
โโใใใฃใฆใใใใ
|
70 |
+
ใใขใผใญใณใฏไธๆตใซ็ฌใฃใใใใฎ๏ฟฝ
|
71 |
+
```
|
72 |
+
|
73 |
+
## Acknowledgements
|
74 |
+
|
75 |
+
This project was possible because of the compute provided by the
|
76 |
+
[TPU Research Cloud](https://sites.research.google/trc/)
|
77 |
+
|
78 |
+
Thanks [EleutherAI](https://eleuther.ai/) for pretraining the GPT-J 6B model.
|
79 |
+
|
80 |
+
Thanks to everyone who contributed to this project!
|
81 |
+
|
82 |
+
- [Finetune](https://github.com/finetuneanon)
|
83 |
+
- [Aero](https://github.com/AeroScripts)
|
84 |
+
- [Kurumuz](https://github.com/kurumuz)
|