Text Generation
Transformers
PyTorch
Safetensors
Japanese
English
gpt_neox
text-generation-inference
tianyuz commited on
Commit
86b9caf
1 Parent(s): a365fbe
Files changed (7) hide show
  1. README.md +143 -0
  2. config.json +27 -0
  3. pytorch_model.bin +3 -0
  4. rinna.png +0 -0
  5. spiece.model +3 -0
  6. spiece.vocab +0 -0
  7. tokenizer_config.json +1 -0
README.md CHANGED
@@ -1,3 +1,146 @@
1
  ---
 
2
  license: mit
 
 
 
 
 
 
3
  ---
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
  ---
2
+ thumbnail: https://github.com/rinnakk/japanese-pretrained-models/blob/master/rinna.png
3
  license: mit
4
+ datasets:
5
+ - Anthropic/hh-rlhf
6
+ language:
7
+ - ja
8
+ - en
9
+ inference: false
10
  ---
11
+
12
+ # bilingual-gpt-neox-4b-instruction-ppo
13
+
14
+ ![rinna-icon](./rinna.png)
15
+
16
+ ---
17
+
18
+ # Overview
19
+ This repository provides an English-Japanese bilingual GPT-NeoX model of 3.8 billion parameters.
20
+
21
+ The model is based on [`rinna/bilingual-gpt-neox-4b-instruction-sft`](https://huggingface.co/rinna/bilingual-gpt-neox-4b-instruction-sft) and has been aligned to serve as an instruction-following conversational agent.
22
+
23
+ * **Model architecture**
24
+
25
+ A 36-layer, 2816-hidden-size transformer-based language model.
26
+
27
+ * **RLHF**
28
+
29
+ Following the [OpenAI InstructGPT paper](https://arxiv.org/abs/2203.02155), **Reinforcement Learning from Human Feedback** (RLHF) has been applied to aligning the model's behaviour with input instructions. Particularly, the model has been trained in two stages, i.e. **Supervised Fine-Tuning** (SFT) and [PPO](https://arxiv.org/abs/1707.06347)-based **Reinforcement Learning** (RL).
30
+ * The first SFT stage produces [`rinna/bilingual-gpt-neox-4b-instruction-sft`](https://huggingface.co/rinna/bilingual-gpt-neox-4b-instruction-sft).
31
+ * The second RL stage produces this model.
32
+
33
+ * **Model Series**
34
+
35
+ | Variant | Link |
36
+ | :-- | :--|
37
+ | Bilingual 4B MiniGPT4 | https://huggingface.co/rinna/bilingual-gpt-neox-4b-minigpt4 |
38
+ | Bilingual 4B PPO | https://huggingface.co/rinna/bilingual-gpt-neox-4b-instruction-ppo |
39
+ | Bilingual 4B SFT | https://huggingface.co/rinna/bilingual-gpt-neox-4b-instruction-sft |
40
+ | Bilingual 4B 8K | https://huggingface.co/rinna/bilingual-gpt-neox-4b-8k |
41
+ | Bilingual 4B | https://huggingface.co/rinna/bilingual-gpt-neox-4b |
42
+ | Japanese 3.6B PPO | https://huggingface.co/rinna/japanese-gpt-neox-3.6b-instruction-ppo |
43
+ | Japanese 3.6B SFT-v2 | https://huggingface.co/rinna/japanese-gpt-neox-3.6b-instruction-sft-v2 |
44
+ | Japanese 3.6B SFT | https://huggingface.co/rinna/japanese-gpt-neox-3.6b-instruction-sft |
45
+ | Japanese 3.6B | https://huggingface.co/rinna/japanese-gpt-neox-3.6b |
46
+
47
+ * **Authors**
48
+
49
+ [Tianyu Zhao](https://huggingface.co/tianyuz) and [Kei Sawada](https://huggingface.co/keisawada)
50
+
51
+ ---
52
+
53
+ # I/O Format
54
+ A special format has been adopted to construct inputs.
55
+ * An input prompt is formatted as a conversation between `ユーザー` and `システム`.
56
+ * Each input utterance consists of (1) its speaker (`"ユーザー"` or `"システム"`), (2) a colon (`":"`), (3) a whitespace (`" "`), and (4) utterance text (e.g. `"世界で一番高い山は?"`).
57
+ * The input prompt should be ended with `"システム: "` to acknowledge the model to generate a response.
58
+ * All the utterances in the input prompt should be separated by a newline `\n`.
59
+
60
+ Following is an example to construct input from a conversation.
61
+ ~~~python
62
+ prompt = [
63
+ {
64
+ "speaker": "ユーザー",
65
+ "text": "Hello, you are an assistant that helps me learn Japanese."
66
+ },
67
+ {
68
+ "speaker": "システム",
69
+ "text": "Sure, what can I do for you?"
70
+ },
71
+ {
72
+ "speaker": "ユーザー",
73
+ "text": "VRはなんですか。"
74
+ }
75
+ ]
76
+ prompt = [
77
+ f"{uttr['speaker']}: {uttr['text']}"
78
+ for uttr in prompt
79
+ ]
80
+ prompt = "\n".join(prompt)
81
+ prompt = (
82
+ prompt
83
+ + "\n"
84
+ + "システム: "
85
+ )
86
+ print(prompt)
87
+ """
88
+ ユーザー: Hello, you are an assistant that helps me learn Japanese.
89
+ システム: Sure, what can I do for you?
90
+ ユーザー: VRはなんですか。
91
+ システム:
92
+ """
93
+ ~~~
94
+
95
+ ---
96
+
97
+ # How to use the model
98
+
99
+ **Notice:** Since the model is **sensitive to decoding hyper-parameters** (e.g. `temperature`, `top_p`, `top_k`, `repetition_penalty`), it is suggested to explore the best setting for your task.
100
+
101
+ ~~~~python
102
+ import torch
103
+ from transformers import AutoTokenizer, AutoModelForCausalLM
104
+
105
+ tokenizer = AutoTokenizer.from_pretrained("rinna/bilingual-gpt-neox-4b-instruction-ppo", use_fast=False)
106
+ model = AutoModelForCausalLM.from_pretrained("rinna/bilingual-gpt-neox-4b-instruction-ppo")
107
+
108
+ if torch.cuda.is_available():
109
+ model = model.to("cuda")
110
+
111
+ token_ids = tokenizer.encode(prompt, add_special_tokens=False, return_tensors="pt")
112
+
113
+ with torch.no_grad():
114
+ output_ids = model.generate(
115
+ token_ids.to(model.device),
116
+ max_new_tokens=512,
117
+ do_sample=True,
118
+ temperature=1.0,
119
+ top_p=0.85,
120
+ pad_token_id=tokenizer.pad_token_id,
121
+ bos_token_id=tokenizer.bos_token_id,
122
+ eos_token_id=tokenizer.eos_token_id
123
+ )
124
+
125
+ output = tokenizer.decode(output_ids.tolist()[0][token_ids.size(1):])
126
+ print(output)
127
+ """VRとはVirtual Realityの略で、仮想現実とも呼ばれます。これは、コンピューターを使用して仮想世界を作り出し、仮想世界上でコンピューターのゲームや仮想世界を体験するための技術です。この技術は、コンピューターやモバイ ルデバイスの進歩に���って、2015年以降、ますます普及しています。VRは、ゲームや仮想世界、その他のアプリケー ションなどのさまざまな分野で、コンピューターと人間の相互作用の新しい方法を提供しています。</s>"""
128
+ ~~~~
129
+
130
+ ---
131
+
132
+ # Tokenization
133
+ The model uses a [sentencepiece](https://github.com/google/sentencepiece)-based tokenizer.
134
+ * The tokenizer has a vocabulary size of 65,536.
135
+ * It uses *byte fallback* to decompose unknown text pieces into UTF-8 byte pieces to avoid producing `<UNK>` tokens.
136
+ * It can recognize *consecutive whitespaces*, *newlines*, and *tabs* to handle structured texts better.
137
+ * We turned off the default behaviour of prepending leading whitespace because it is not beneficial for processing Japanese.
138
+ * Specifically, single whitespace is always processed as one token so that any English word won't have a preceding whitespace like in many other tokenizers (e.g. `_Hello`).
139
+ * This decision trades the English processing efficiency for a unified way to treat whitespaces.
140
+ * It leads to a significantly lower loss of next token prediction on English data because whitespaces are easy to predict.
141
+ * **Don't forget to set `use_fast=False` to make the above features function correctly.**
142
+
143
+ ---
144
+
145
+ # Licenese
146
+ [The MIT license](https://opensource.org/licenses/MIT)
config.json ADDED
@@ -0,0 +1,27 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "architectures": [
3
+ "GPTNeoXForCausalLM"
4
+ ],
5
+ "attention_dropout": 0.1,
6
+ "bos_token_id": 2,
7
+ "classifier_dropout": 0.1,
8
+ "eos_token_id": 3,
9
+ "hidden_act": "gelu",
10
+ "hidden_dropout": 0.1,
11
+ "hidden_size": 2816,
12
+ "initializer_range": 0.02,
13
+ "intermediate_size": 11264,
14
+ "layer_norm_eps": 1e-05,
15
+ "max_position_embeddings": 2048,
16
+ "model_type": "gpt_neox",
17
+ "num_attention_heads": 22,
18
+ "num_hidden_layers": 36,
19
+ "rope_scaling": null,
20
+ "rotary_emb_base": 10000,
21
+ "rotary_pct": 1.0,
22
+ "tie_word_embeddings": false,
23
+ "torch_dtype": "float16",
24
+ "use_cache": true,
25
+ "use_parallel_residual": false,
26
+ "vocab_size": 65536
27
+ }
pytorch_model.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:445c7189f5a6451285e1711b05c8c65778f98d19450b93e1d3c4306086c3faa7
3
+ size 7775149229
rinna.png ADDED
spiece.model ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:85a0205d37a98bb3b97cf4ca3f507c78873cf8f6cefa3b51d8d6a15006dc889d
3
+ size 1341798
spiece.vocab ADDED
The diff for this file is too large to render. See raw diff
 
tokenizer_config.json ADDED
@@ -0,0 +1 @@
 
 
1
+ {"eos_token": "</s>", "unk_token": "[UNK]", "pad_token": "[PAD]", "extra_ids": 0, "additional_special_tokens": [], "sp_model_kwargs": {}, "bos_token": "<s>", "cls_token": "[CLS]", "sep_token": "[SEP]", "mask_token": "[MASK]", "do_lower_case": false, "tokenizer_class": "T5Tokenizer"}