losyer8 commited on
Commit
d3e1c81
1 Parent(s): c156c5e

Create README.md

Browse files
Files changed (1) hide show
  1. README.md +154 -0
README.md ADDED
@@ -0,0 +1,154 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: apache-2.0
3
+ language:
4
+ - en
5
+ - ja
6
+ programming_language:
7
+ - C
8
+ - C++
9
+ - C#
10
+ - Go
11
+ - Java
12
+ - JavaScript
13
+ - Lua
14
+ - PHP
15
+ - Python
16
+ - Ruby
17
+ - Rust
18
+ - Scala
19
+ - TypeScript
20
+ library_name: transformers
21
+ pipeline_tag: text-generation
22
+ inference: false
23
+ ---
24
+ # llm-jp-13b-instruct-full-jaster-v1.0
25
+
26
+ This repository provides large language models developed by [LLM-jp](https://llm-jp.nii.ac.jp/), a collaborative project launched in Japan.
27
+
28
+ | Model Variant |
29
+ | :--- |
30
+ |**Instruction models**|
31
+ | [llm-jp-13b-instruct-full-jaster-v1.0](https://huggingface.co/llm-jp/llm-jp-13b-instruct-full-jaster-v1.0) |
32
+ | [llm-jp-13b-instruct-full-jaster-dolly-oasst-v1.0](https://huggingface.co/llm-jp/llm-jp-13b-instruct-full-jaster-dolly-oasst-v1.0) |
33
+ | [llm-jp-13b-instruct-full-dolly-oasst-v1.0](https://huggingface.co/llm-jp/llm-jp-13b-instruct-full-dolly-oasst-v1.0) |
34
+ | [llm-jp-13b-instruct-lora-jaster-v1.0](https://huggingface.co/llm-jp/llm-jp-13b-instruct-lora-jaster-v1.0) |
35
+ | [llm-jp-13b-instruct-lora-jaster-dolly-oasst-v1.0](https://huggingface.co/llm-jp/llm-jp-13b-instruct-lora-jaster-dolly-oasst-v1.0) |
36
+ | [llm-jp-13b-instruct-lora-dolly-oasst-v1.0](https://huggingface.co/llm-jp/llm-jp-13b-instruct-lora-dolly-oasst-v1.0) |
37
+
38
+
39
+ | |
40
+ | :--- |
41
+ |**Pre-trained models**|
42
+ | [llm-jp-13b-v1.0](https://huggingface.co/llm-jp/llm-jp-13b-v1.0) |
43
+ | [llm-jp-1.3b-v1.0](https://huggingface.co/llm-jp/llm-jp-1.3b-v1.0) |
44
+ Checkpoints format: `transformers` (Megatron-DeepSpeed format available [here](https://huggingface.co/llm-jp/llm-jp-13b-v1.0-mdsfmt))
45
+
46
+
47
+ ## Required Libraries and Their Versions
48
+
49
+ - torch>=2.0.0
50
+ - transformers>=4.34.0
51
+ - tokenizers>=0.14.0
52
+
53
+ ## Usage
54
+
55
+ ```python
56
+ import torch
57
+ from transformers import AutoTokenizer, AutoModelForCausalLM
58
+ tokenizer = AutoTokenizer.from_pretrained("llm-jp/[Model_Name]")
59
+ model = AutoModelForCausalLM.from_pretrained("llm-jp/[Model_Name]", torch_dtype=torch.float16)
60
+ text = "自然言語処理とは何か"
61
+ text = text + "### 回答:"
62
+ tokenized_input = tokenizer.encode(text, add_special_tokens=False, return_tensors="pt").to(model.device)
63
+ with torch.no_grad():
64
+ output = model.generate(
65
+ tokenized_input,
66
+ max_new_tokens=100,
67
+ do_sample=True,
68
+ top_p=0.95,
69
+ temperature=0.7,
70
+ )[0]
71
+ print(tokenizer.decode(output))
72
+ ```
73
+
74
+
75
+ ## Model Details
76
+
77
+ - **Model type:** Transformer-based Language Model
78
+ - **Total seen tokens:** 270B+
79
+
80
+ |Model|Params|Layers|Hidden size|Heads|Context length|
81
+ |:---:|:---:|:---:|:---:|:---:|:---:|
82
+ |13b model|13b|40|5120|40|2048|
83
+ |1.3b model|1.3b|24|2048|16|2048|
84
+
85
+
86
+ ## Training
87
+
88
+ - **Pre-training:**
89
+ - **Hardware:** 96 A100 40GB GPUs ([mdx cluster](https://mdx.jp/en/))
90
+ - **Software:** Megatron-DeepSpeed
91
+
92
+ - **Instruction tuning:**
93
+ - **Hardware:** 8 A100 40GB GPUs ([mdx cluster](https://mdx.jp/en/))
94
+ - **Software:** [TRL](https://github.com/huggingface/trl), [PEFT](https://github.com/huggingface/peft), and [DeepSpeed](https://github.com/microsoft/DeepSpeed)
95
+
96
+ ## Tokenizer
97
+ The tokenizer of this model is based on [huggingface/tokenizers](https://github.com/huggingface/tokenizers) Unigram byte-fallback model.
98
+ The vocab entries were converted from [`llm-jp-tokenizer v2.1 (50k)`](https://github.com/llm-jp/llm-jp-tokenizer/releases/tag/v2.1).
99
+ Please refer to [README.md](https://github.com/llm-jp/llm-jp-tokenizer) of `llm-ja-tokenizer` for the details of vocab constuction steps.
100
+ - **Model:** Hugging Face Fast Tokenizer using Unigram byte-fallback model which requires `tokenizers>=0.14.0`
101
+ - **Training algorithm:** SentencePiece Unigram byte-fallback
102
+ - **Training data:** A subset of the datasets for model pre-training
103
+ - **Vocabulary size:** 50,570 (mixed vocabulary of Japanese, English, and source code)
104
+
105
+
106
+ ## Datasets
107
+
108
+ ### Pre-training
109
+
110
+ The models have been pre-trained on approximately 287.5B tokens, sourced from a blend of the following datasets.
111
+
112
+ | Language | Dataset | Tokens|
113
+ |:---:|:---:|:---:|
114
+ |Japanese|[Wikipedia](https://huggingface.co/datasets/wikipedia)|1.5B
115
+ ||[mC4](https://huggingface.co/datasets/mc4)|136B
116
+ |English|[Wikipedia](https://huggingface.co/datasets/wikipedia)|5B
117
+ ||[The Pile](https://huggingface.co/datasets/EleutherAI/pile)|135B
118
+ |Codes|[The Stack](https://huggingface.co/datasets/bigcode/the-stack)|10B
119
+
120
+ Pretraining was done by 10-hold shards that consists approx. 27-28B tokens. We further finalized the pretraining with additional cleaned 27B tokens data.
121
+
122
+ ### Instruction tuning
123
+
124
+ The models have been fine-tuned on the following datasets.
125
+
126
+ | Language | Dataset | description |
127
+ |:---|:---:|:---:|
128
+ |Japanese|[jaster](https://github.com/llm-jp/llm-jp-eval)| An automatically transformed data from the existing Japanese NLP datasets |
129
+ ||[databricks-dolly-15k](https://huggingface.co/datasets/databricks/databricks-dolly-15k)| A translated one by DeepL in LLM-jp |
130
+ ||[OpenAssistant Conversations Dataset](https://huggingface.co/datasets/OpenAssistant/oasst1)| A translated one by DeepL in LLM-jp |
131
+
132
+
133
+ ## Evaluation
134
+ You can view the evaluation results of several LLMs on this [leaderboard](http://wandb.me/llm-jp-leaderboard). We used [llm-jp-eval](https://github.com/llm-jp/llm-jp-eval) for the evaluation.
135
+
136
+ ## Risks and Limitations
137
+
138
+ The models released here are still in the early stages of our research and development and have not been tuned to ensure outputs align with human intent and safety considerations.
139
+
140
+
141
+ ## Send Questions to
142
+
143
+ llm-jp(at)nii.ac.jp
144
+
145
+
146
+ ## License
147
+
148
+ [Apache License, Version 2.0](https://www.apache.org/licenses/LICENSE-2.0)
149
+
150
+
151
+ ## Model Card Authors
152
+ *The names are listed in alphabetical order.*
153
+
154
+ Namgi Han, Hirokazu Kiyomaru, Hiroshi Matsuda, Shota Sasaki, Shuhei Kurita, Taishi Nakamura, Takumi Okamoto.