Update README.md
Browse files
README.md
CHANGED
@@ -1,83 +1,71 @@
|
|
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 |
-
datasets:
|
24 |
-
- databricks/databricks-dolly-15k
|
25 |
-
- llm-jp/databricks-dolly-15k-ja
|
26 |
-
- llm-jp/oasst1-21k-en
|
27 |
-
- llm-jp/oasst1-21k-ja
|
28 |
---
|
29 |
-
|
30 |
# llm-jp-13b-instruct-full-dolly-ichikara_004_001_single-oasst-oasst2-v2.0
|
31 |
|
32 |
This repository provides large language models developed by [LLM-jp](https://llm-jp.nii.ac.jp/), a collaborative project launched in Japan.
|
33 |
|
34 |
| Model Variant |
|
35 |
| :--- |
|
36 |
-
|**Instruction models
|
37 |
-
| [llm-jp-13b-
|
38 |
-
| [llm-jp-13b-instruct-full-
|
39 |
-
| [llm-jp-13b-instruct-
|
40 |
-
|**Instruction models ver1.0**|
|
41 |
-
| [llm-jp-13b-instruct-full-jaster-v1.0](https://huggingface.co/llm-jp/llm-jp-13b-instruct-full-jaster-v1.0) |
|
42 |
-
| [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) |
|
43 |
-
| [llm-jp-13b-instruct-full-dolly-oasst-v1.0](https://huggingface.co/llm-jp/llm-jp-13b-instruct-full-dolly-oasst-v1.0) |
|
44 |
-
| [llm-jp-13b-instruct-lora-jaster-v1.0](https://huggingface.co/llm-jp/llm-jp-13b-instruct-lora-jaster-v1.0) |
|
45 |
-
| [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) |
|
46 |
-
| [llm-jp-13b-instruct-lora-dolly-oasst-v1.0](https://huggingface.co/llm-jp/llm-jp-13b-instruct-lora-dolly-oasst-v1.0) |
|
47 |
|
48 |
|
49 |
| |
|
50 |
| :--- |
|
51 |
|**Pre-trained models**|
|
52 |
-
| [llm-jp-13b-
|
53 |
-
|
54 |
-
Checkpoints format: Hugging Face Transformers
|
55 |
|
56 |
|
57 |
## Required Libraries and Their Versions
|
58 |
|
59 |
-
- torch>=2.
|
60 |
-
- transformers>=4.
|
61 |
-
- tokenizers>=0.
|
62 |
-
- accelerate
|
|
|
63 |
|
64 |
## Usage
|
65 |
|
66 |
```python
|
67 |
import torch
|
68 |
from transformers import AutoTokenizer, AutoModelForCausalLM
|
69 |
-
tokenizer = AutoTokenizer.from_pretrained("llm-jp/llm-jp-13b-instruct-full-
|
70 |
-
model = AutoModelForCausalLM.from_pretrained("llm-jp/llm-jp-13b-instruct-full-
|
71 |
-
text = "
|
72 |
tokenized_input = tokenizer.encode(text, add_special_tokens=False, return_tensors="pt").to(model.device)
|
73 |
with torch.no_grad():
|
74 |
output = model.generate(
|
75 |
tokenized_input,
|
76 |
-
max_new_tokens=
|
77 |
do_sample=True,
|
78 |
top_p=0.95,
|
79 |
temperature=0.7,
|
80 |
-
repetition_penalty=1.
|
81 |
)[0]
|
82 |
print(tokenizer.decode(output))
|
83 |
```
|
@@ -86,32 +74,34 @@ print(tokenizer.decode(output))
|
|
86 |
## Model Details
|
87 |
|
88 |
- **Model type:** Transformer-based Language Model
|
89 |
-
- **Total seen tokens:**
|
90 |
|
91 |
|Model|Params|Layers|Hidden size|Heads|Context length|
|
92 |
|:---:|:---:|:---:|:---:|:---:|:---:|
|
93 |
-
|13b model|13b|40|5120|40|
|
94 |
-
|1.3b model|1.3b|24|2048|16|2048|
|
95 |
|
96 |
|
97 |
## Training
|
98 |
|
99 |
- **Pre-training:**
|
100 |
-
- **Hardware:**
|
101 |
-
- **Software:** Megatron-
|
102 |
|
103 |
- **Instruction tuning:**
|
104 |
- **Hardware:** 8 A100 40GB GPUs ([mdx cluster](https://mdx.jp/en/))
|
105 |
- **Software:** [TRL](https://github.com/huggingface/trl), [PEFT](https://github.com/huggingface/peft), and [DeepSpeed](https://github.com/microsoft/DeepSpeed)
|
106 |
|
107 |
## Tokenizer
|
|
|
108 |
The tokenizer of this model is based on [huggingface/tokenizers](https://github.com/huggingface/tokenizers) Unigram byte-fallback model.
|
109 |
-
The vocabulary entries were converted from [`llm-jp-tokenizer v2.
|
110 |
-
Please refer to [README.md](https://github.com/llm-jp/llm-jp-tokenizer) of `llm-ja-tokenizer` for details on the vocabulary construction procedure.
|
|
|
111 |
- **Model:** Hugging Face Fast Tokenizer using Unigram byte-fallback model which requires `tokenizers>=0.14.0`
|
112 |
-
- **Training algorithm:** SentencePiece Unigram byte-fallback
|
113 |
- **Training data:** A subset of the datasets for model pre-training
|
114 |
-
- **Vocabulary size:**
|
|
|
115 |
|
116 |
|
117 |
## Datasets
|
@@ -121,33 +111,33 @@ Please refer to [README.md](https://github.com/llm-jp/llm-jp-tokenizer) of `llm-
|
|
121 |
The models have been pre-trained using a blend of the following datasets.
|
122 |
|
123 |
| Language | Dataset | Tokens|
|
124 |
-
|
125 |
-
|Japanese|[Wikipedia](https://huggingface.co/datasets/wikipedia)|1.
|
126 |
-
||[
|
127 |
-
|English|[Wikipedia](https://huggingface.co/datasets/wikipedia)|
|
128 |
-
||[The Pile](https://huggingface.co/datasets/EleutherAI/pile)|
|
129 |
-
|Codes|[The Stack](https://huggingface.co/datasets/bigcode/the-stack)|
|
130 |
-
|
131 |
-
The pre-training was continuously conducted using a total of 10 folds of non-overlapping data, each consisting of approximately 27-28B tokens.
|
132 |
-
We finalized the pre-training with additional (potentially) high-quality 27B tokens data obtained from the identical source datasets listed above used for the 10-fold data.
|
133 |
|
134 |
### Instruction tuning
|
135 |
|
136 |
The models have been fine-tuned on the following datasets.
|
137 |
|
138 |
| Language | Dataset | description |
|
139 |
-
|
140 |
-
|Japanese|[
|
141 |
-
|
|
142 |
-
|
|
143 |
-
|
|
144 |
-
|
|
145 |
-
|
|
146 |
-
|
|
147 |
-
|
148 |
|
149 |
## Evaluation
|
150 |
-
|
|
|
|
|
|
|
|
|
151 |
|
152 |
## Risks and Limitations
|
153 |
|
@@ -165,6 +155,7 @@ llm-jp(at)nii.ac.jp
|
|
165 |
|
166 |
|
167 |
## Model Card Authors
|
|
|
168 |
*The names are listed in alphabetical order.*
|
169 |
|
170 |
-
|
|
|
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-dolly-ichikara_004_001_single-oasst-oasst2-v2.0
|
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-dolly-ichikara_004_001_single-oasst-oasst2-v2.0](https://huggingface.co/llm-jp/llm-jp-13b-instruct-full-dolly-ichikara_004_001_single-oasst-oasst2-v2.0) |
|
32 |
+
| [llm-jp-13b-instruct-full-ac_001-dolly-ichikara_004_001_single-oasst-oasst2-v2.0](https://huggingface.co/llm-jp/llm-jp-13b-instruct-full-ac_001-dolly-ichikara_004_001_single-oasst-oasst2-v2.0) |
|
33 |
+
| [llm-jp-13b-instruct-full-ac_001_16x-dolly-ichikara_004_001_single-oasst-oasst2-v2.0](https://huggingface.co/llm-jp/llm-jp-13b-instruct-full-ac_001_16x-dolly-ichikara_004_001_single-oasst-oasst2-v2.0) |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
34 |
|
35 |
|
36 |
| |
|
37 |
| :--- |
|
38 |
|**Pre-trained models**|
|
39 |
+
| [llm-jp-13b-v2.0](https://huggingface.co/llm-jp/llm-jp-13b-v2.0) |
|
40 |
+
|
41 |
+
Checkpoints format: Hugging Face Transformers
|
42 |
|
43 |
|
44 |
## Required Libraries and Their Versions
|
45 |
|
46 |
+
- torch>=2.3.0
|
47 |
+
- transformers>=4.40.1
|
48 |
+
- tokenizers>=0.19.1
|
49 |
+
- accelerate>=0.29.3
|
50 |
+
- flash-attn>=2.5.8
|
51 |
|
52 |
## Usage
|
53 |
|
54 |
```python
|
55 |
import torch
|
56 |
from transformers import AutoTokenizer, AutoModelForCausalLM
|
57 |
+
tokenizer = AutoTokenizer.from_pretrained("llm-jp/llm-jp-13b-instruct-full-dolly-ichikara_004_001_single-oasst-oasst2-v2.0")
|
58 |
+
model = AutoModelForCausalLM.from_pretrained("llm-jp/llm-jp-13b-instruct-full-dolly-ichikara_004_001_single-oasst-oasst2-v2.0", device_map="auto", torch_dtype=torch.float16)
|
59 |
+
text = "自然言語処理とは何か"
|
60 |
tokenized_input = tokenizer.encode(text, add_special_tokens=False, return_tensors="pt").to(model.device)
|
61 |
with torch.no_grad():
|
62 |
output = model.generate(
|
63 |
tokenized_input,
|
64 |
+
max_new_tokens=100,
|
65 |
do_sample=True,
|
66 |
top_p=0.95,
|
67 |
temperature=0.7,
|
68 |
+
repetition_penalty=1.05,
|
69 |
)[0]
|
70 |
print(tokenizer.decode(output))
|
71 |
```
|
|
|
74 |
## Model Details
|
75 |
|
76 |
- **Model type:** Transformer-based Language Model
|
77 |
+
- **Total seen tokens:** 256B
|
78 |
|
79 |
|Model|Params|Layers|Hidden size|Heads|Context length|
|
80 |
|:---:|:---:|:---:|:---:|:---:|:---:|
|
81 |
+
|13b model|13b|40|5120|40|4096|
|
|
|
82 |
|
83 |
|
84 |
## Training
|
85 |
|
86 |
- **Pre-training:**
|
87 |
+
- **Hardware:** 128 A100 40GB GPUs ([mdx cluster](https://mdx.jp/en/))
|
88 |
+
- **Software:** Megatron-LM
|
89 |
|
90 |
- **Instruction tuning:**
|
91 |
- **Hardware:** 8 A100 40GB GPUs ([mdx cluster](https://mdx.jp/en/))
|
92 |
- **Software:** [TRL](https://github.com/huggingface/trl), [PEFT](https://github.com/huggingface/peft), and [DeepSpeed](https://github.com/microsoft/DeepSpeed)
|
93 |
|
94 |
## Tokenizer
|
95 |
+
|
96 |
The tokenizer of this model is based on [huggingface/tokenizers](https://github.com/huggingface/tokenizers) Unigram byte-fallback model.
|
97 |
+
The vocabulary entries were converted from [`llm-jp-tokenizer v2.2 (100k: code20K_en40K_ja60K.ver2.2)`](https://github.com/llm-jp/llm-jp-tokenizer/releases/tag/v2.2).
|
98 |
+
Please refer to [README.md](https://github.com/llm-jp/llm-jp-tokenizer) of `llm-ja-tokenizer` for details on the vocabulary construction procedure (the pure SentencePiece training does not reproduce our vocabulary).
|
99 |
+
|
100 |
- **Model:** Hugging Face Fast Tokenizer using Unigram byte-fallback model which requires `tokenizers>=0.14.0`
|
101 |
+
- **Training algorithm:** Marging Code/English/Japanese vocabularies constructed with SentencePiece Unigram byte-fallback and reestimating scores with the EM-algorithm.
|
102 |
- **Training data:** A subset of the datasets for model pre-training
|
103 |
+
- **Vocabulary size:** 96,867 (mixed vocabulary of Japanese, English, and source code)
|
104 |
+
- The acutal size of vocabulary in the pretrained model is 97,024 due to round-up to multiples of 256.
|
105 |
|
106 |
|
107 |
## Datasets
|
|
|
111 |
The models have been pre-trained using a blend of the following datasets.
|
112 |
|
113 |
| Language | Dataset | Tokens|
|
114 |
+
|:---|:---|---:|
|
115 |
+
|Japanese|[Wikipedia](https://huggingface.co/datasets/wikipedia)|1.4B
|
116 |
+
||[Common Crawl](https://gitlab.llm-jp.nii.ac.jp/datasets/llm-jp-corpus-v2)|130.7B
|
117 |
+
|English|[Wikipedia](https://huggingface.co/datasets/wikipedia)|4.7B
|
118 |
+
||[The Pile](https://huggingface.co/datasets/EleutherAI/pile)|110.3B
|
119 |
+
|Codes|[The Stack](https://huggingface.co/datasets/bigcode/the-stack)|8.7B
|
|
|
|
|
|
|
120 |
|
121 |
### Instruction tuning
|
122 |
|
123 |
The models have been fine-tuned on the following datasets.
|
124 |
|
125 |
| Language | Dataset | description |
|
126 |
+
|:---|:---|:---|
|
127 |
+
|Japanese|[ichikara-instruction-004-001](https://liat-aip.sakura.ne.jp/wp/llm%e3%81%ae%e3%81%9f%e3%82%81%e3%81%ae%e6%97%a5%e6%9c%ac%e8%aa%9e%e3%82%a4%e3%83%b3%e3%82%b9%e3%83%88%e3%83%a9%e3%82%af%e3%82%b7%e3%83%a7%e3%83%b3%e3%83%87%e3%83%bc%e3%82%bf%e4%bd%9c%e6%88%90/llm%e3%81%ae%e3%81%9f%e3%82%81%e3%81%ae%e6%97%a5%e6%9c%ac%e8%aa%9e%e3%82%a4%e3%83%b3%e3%82%b9%e3%83%88%e3%83%a9%e3%82%af%e3%82%b7%e3%83%a7%e3%83%b3%e3%83%87%e3%83%bc%e3%82%bf-%e5%85%ac%e9%96%8b/)| A manually constructed Japanese instruction dataset |
|
128 |
+
| |[answer-carefully-001](https://liat-aip.sakura.ne.jp/wp/answercarefully-dataset/)| A manually constructed Japanese instruction dataset focusing on LLMs' safety |
|
129 |
+
| |[databricks-dolly-15k-ja](https://huggingface.co/datasets/llm-jp/databricks-dolly-15k-ja)| [databricks-dolly-15k](https://huggingface.co/datasets/databricks/databricks-dolly-15k) translated into Japanese using DeepL |
|
130 |
+
| |[oasst1-21k-ja](https://huggingface.co/datasets/llm-jp/oasst1-21k-ja)| A subset of [oasst1](https://huggingface.co/datasets/OpenAssistant/oasst1) translated into Japanese using DeepL |
|
131 |
+
| |[oasst2-33k-ja](https://huggingface.co/datasets/llm-jp/oasst2-33k-ja)| A subset of [oasst2](https://huggingface.co/datasets/OpenAssistant/oasst2) translated into Japanese using DeepL |
|
132 |
+
|English |[oasst1-21k-en](https://huggingface.co/datasets/llm-jp/oasst1-21k-en)| A subset of [oasst1](https://huggingface.co/datasets/OpenAssistant/oasst1) |
|
133 |
+
| |[oasst2-33k-en](https://huggingface.co/datasets/llm-jp/oasst2-33k-en)| A subset of [oasst2](https://huggingface.co/datasets/OpenAssistant/oasst2) |
|
|
|
134 |
|
135 |
## Evaluation
|
136 |
+
|
137 |
+
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) (v1.3.0) for the evaluation.
|
138 |
+
|
139 |
+
Besides, we used LLM-as-a-judge frameworks, [Japanese Vicuna QA Benchmark](https://github.com/ku-nlp/ja-vicuna-qa-benchmark/) and [Japanese MT Bench](https://github.com/Stability-AI/FastChat/tree/jp-stable/fastchat/llm_judge), for evaluation.
|
140 |
+
For details, please refer to [our technical blog](https://llm-jp.nii.ac.jp/blog/2024/04/30/v2.0-release.html) (in Japanese).
|
141 |
|
142 |
## Risks and Limitations
|
143 |
|
|
|
155 |
|
156 |
|
157 |
## Model Card Authors
|
158 |
+
|
159 |
*The names are listed in alphabetical order.*
|
160 |
|
161 |
+
Namgi Han, Tatsuya Hiraoka, Hirokazu Kiyomaru, Takashi Kodama, and Hiroshi Matsuda.
|