File size: 4,260 Bytes
2d9465d
 
6c5fda4
3a5713e
2d9465d
 
 
 
 
 
 
 
 
 
57d1fd4
2d9465d
 
 
3a5713e
2d9465d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
74c1cb8
2d9465d
 
 
5b4dc79
 
2d9465d
 
 
 
 
 
 
 
 
 
 
 
 
 
fcbe07a
2d9465d
 
 
 
 
 
 
 
9ac3417
2d9465d
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
---
language:
- ja
- en
tags:
- pytorch
- causal-lm
license: apache-2.0

---

# Genji-JP 6B

Please check our blog post for more details, samples, evaluations and more:
[Blogpost](https://blog.novelai.net/data-efficient-language-transfer-with-gpt-j-45daedaaf35a)

## Model Description

Genji-JP 6B is a model finetuned on our Japanese storytelling dataset based on EleutherAI's GPT-J 6B model. This particular model is trained on Japanese web novels.

| Hyperparameter    | Value  | 
|-------------------|--------|
| n_parameters      | 6,053,381,344 |
| n_layers          | 28*    |
| d_model           | 4,096  |
| d_ff              | 16,384 |
| n_heads           | 16     |
| d_head            | 256    |
| n_ctx             | 2,048  |
| n_vocab           | 50,400 (same tokenizer as GPT-2/3)  |
| position encoding | [Rotary position encodings (RoPE)](https://arxiv.org/abs/2104.09864) |
| RoPE dimensions   | [64](https://github.com/kingoflolz/mesh-transformer-jax/blob/f2aa66e0925de6593dcbb70e72399b97b4130482/mesh_transformer/layers.py#L223) |

`*` each layer consists of one feedforward block and one self attention block

The model consists of 28 layers with a model dimension of 4096, and a feedforward dimension of 16384. The model
dimension is split into 16 heads, each with a dimension of 256. Rotary position encodings (RoPE) was applied to 64
dimensions of each head. The model is trained with a tokenization vocabulary of 50257, using the same set of BPEs as
GPT-2/GPT-3.

## Training data

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 our Japanese storytelling dataset. Check our blog post for more details.

### How to use

```
from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

tokenizer = AutoTokenizer.from_pretrained("EleutherAI/gpt-j-6B")
model = AutoModelForCausalLM.from_pretrained("NovelAI/genji-jp", torch_dtype=torch.float16, low_cpu_mem_usage=True).eval().cuda()
text = '''ใ‚ใ‚‰ใ™ใ˜๏ผšใ‚ใชใŸใฏ็•ฐไธ–็•Œใซ่ปข็”Ÿใ—ใฆใ—ใพใ„ใพใ—ใŸใ€‚ๅ‹‡่€…ใจใชใฃใฆใ€ไปฒ้–“ใ‚’ไฝœใ‚Šใ€็•ฐไธ–็•Œใ‚’ๅ†’้™บใ—ใ‚ˆใ†๏ผ
***
่ปข็”Ÿใ™ใ‚‹ใจใ€ใ‚ใ‚‹่ƒฝๅŠ›ใ‚’ๆ‰‹ใซๅ…ฅใ‚Œใฆใ„ใŸใ€‚ใใ‚Œใฏใ€'''

tokens = tokenizer(text, return_tensors="pt").input_ids
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)
last_tokens = generated_tokens[0]
generated_text = tokenizer.decode(last_tokens).replace("๏ฟฝ", "")
print("Generation:\n" + generated_text)
```
When run, produces output like this:
```
Generation:
ใ‚ใ‚‰ใ™ใ˜๏ผšใ‚ใชใŸใฏ็•ฐไธ–็•Œใซ่ปข็”Ÿใ—ใฆใ—ใพใ„ใพใ—ใŸใ€‚ๅ‹‡่€…ใจใชใฃใฆใ€ไปฒ้–“ใ‚’ไฝœใ‚Šใ€็•ฐไธ–็•Œใ‚’ๅ†’้™บใ—ใ‚ˆใ†๏ผ
***
่ปข็”Ÿใ™ใ‚‹ใจใ€ใ‚ใ‚‹่ƒฝๅŠ›ใ‚’ๆ‰‹ใซๅ…ฅใ‚Œใฆใ„ใŸใ€‚ใใ‚Œใฏใ€ใ€Žไบˆ็Ÿฅใ€ใ ใ€‚้ŽๅŽปใ‹ใ‚‰ๆœชๆฅใฎใ“ใจใ‚’ใ€่ชฐใ‚‚็Ÿฅใ‚‰ใชใ„ๅ‡บๆฅไบ‹ใ‚‚ๅซใ‚ใฆ่ฆ‹้€šใ™ใ“ใจใŒๅ‡บๆฅใ‚‹ใ€‚
ๆ‚ช้ญ”ใฎๆฌ ็‰‡ใจๅ‘ผใฐใ‚Œใ‚‹ๅฐใ•ใช็ตๆ™ถใ‚’ๅ–ใ‚Š่พผใ‚“ใงใ€ไฝฟๅฝนใ™ใ‚‹ใ“ใจใŒๅ‡บๆฅใ‚‹ใ€‚ไบบใ‚’ๆƒนใใคใ‘ใ€ๅ •่ฝใ•ใ›ใ‚‹ใ€‚ไฝ•ใ‚ˆใ‚Šใ€ไฟบใฏ็”ทใชใ‚“ใฆๅฑ…ใชใ‹ใฃใŸใ—ใ€ๅฅณใซ่ˆˆๅ‘ณใ‚‚ใชใ„ใ€‚โ€ฆโ€ฆใใ‚“ใชใ‚ฏใ‚บใฎ็‰‡ๆฃ’ใ‚’ๆ‹…ใŽไธŠใ’ใ‚‹ๅฅดใŒๅคšใใชใ‚‹ใจๆ€ใ†ใจใ€ใกใ‚‡ใฃใจ่‹ฆใ—ใ„ใ€‚
ใ ใŒใ€ไธ€้ƒจใฎไบบ้–“ใซใฏๅ”ๅŠ›่€…ใ‚’ๅพ—ใ‚‹ใ“ใจใŒๅ‡บๆฅใ‚‹ใ€‚็›ฎ็ซ‹ใŸใชใ„่ก—ใซใ‚ใ‚‹ๅฏบใฎไธญใงใ€ๅธธใซๅฎถใซๅผ•ใใ“ใ‚‚ใฃใฆใ„ใ‚‹่€ไบบใ€‚ใใ‚“ใชใƒคใƒ„ใฎ้ญ‚ใ‚’ใ‚ณใƒณใƒˆใƒญใƒผใƒซใ™ใ‚‹ใ“ใจใŒๅ‡บๆฅใ‚‹ใฎใ ใ€‚ไพฟๅˆฉใช่ƒฝๅŠ›ใ ใ€‚ใ—ใ‹ใ—ใ€่ฃๅˆ‡ใ‚Š่€…ใฏๅคงๅ‹ขใ„ใ‚‹ใ€‚ๆฐ—ใ‚’ๆŠœใ‘ใฐใ€็‹‚ใ†ใ€‚ใ ใ‹ใ‚‰ๆณจๆ„ใŒๅฟ…่ฆใ ใ€‚
โ€•โ€•ใ€Œใ‚„ใฃใฆใ‚„ใ‚‹ใ‚ˆใ€
ใ€€ใ‚ขใƒผใƒญใƒณใฏไธๆ•ตใซ็ฌ‘ใฃใŸใ€‚ใ“ใฎ
```

## Acknowledgements

This project was possible because of the compute provided by the
[TPU Research Cloud](https://sites.research.google/trc/)

Thanks [EleutherAI](https://eleuther.ai/) for pretraining the GPT-J 6B model.

Thanks to everyone who contributed to this project!

- [Finetune](https://github.com/finetuneanon)
- [Aero](https://github.com/AeroScripts)
- [Kurumuz](https://github.com/kurumuz)