Create README.md
Browse files
README.md
ADDED
@@ -0,0 +1,219 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
---
|
2 |
+
language:
|
3 |
+
- ko
|
4 |
+
- en
|
5 |
+
license: other
|
6 |
+
library_name: transformers
|
7 |
+
license_name: gemma-terms-of-use
|
8 |
+
license_link: https://ai.google.dev/gemma/terms
|
9 |
+
pipeline_tag: text-generation
|
10 |
+
tags:
|
11 |
+
- pytorch
|
12 |
+
---
|
13 |
+
|
14 |
+
# Gemma-Ko
|
15 |
+
|
16 |
+
> Update @ 2024.03.26: First release of Gemma-Ko 2B model
|
17 |
+
|
18 |
+
**Original Gemma Model Page**: [Gemma](https://ai.google.dev/gemma/docs)
|
19 |
+
|
20 |
+
This model card corresponds to the 7B base version of the **Gemma-Ko** model.
|
21 |
+
|
22 |
+
**Resources and Technical Documentation**:
|
23 |
+
|
24 |
+
* [Original Google's Gemma-2B](https://huggingface.co/google/gemma-2b)
|
25 |
+
* [Training Code @ Github: Gemma-EasyLM](https://github.com/Beomi/Gemma-EasyLM)
|
26 |
+
|
27 |
+
**Terms of Use**: [Terms](https://www.kaggle.com/models/google/gemma/license/consent)
|
28 |
+
|
29 |
+
**Citation**
|
30 |
+
|
31 |
+
```bibtex
|
32 |
+
@misc {gemma_ko_7b,
|
33 |
+
author = { {Junbum Lee, Taekyoon Choi} },
|
34 |
+
title = { gemma-ko-7b },
|
35 |
+
year = 2024,
|
36 |
+
url = { https://huggingface.co/beomi/gemma-ko-7b },
|
37 |
+
doi = { 10.57967/hf/1859 },
|
38 |
+
publisher = { Hugging Face }
|
39 |
+
}
|
40 |
+
```
|
41 |
+
|
42 |
+
**Model Developers**: Junbum Lee (Beomi) & Taekyoon Choi (Taekyoon)
|
43 |
+
|
44 |
+
## Model Information
|
45 |
+
|
46 |
+
Summary description and brief definition of inputs and outputs.
|
47 |
+
|
48 |
+
### Description
|
49 |
+
|
50 |
+
Gemma is a family of lightweight, state-of-the-art open models from Google,
|
51 |
+
built from the same research and technology used to create the Gemini models.
|
52 |
+
They are text-to-text, decoder-only large language models, available in English,
|
53 |
+
with open weights, pre-trained variants, and instruction-tuned variants. Gemma
|
54 |
+
models are well-suited for a variety of text generation tasks, including
|
55 |
+
question answering, summarization, and reasoning. Their relatively small size
|
56 |
+
makes it possible to deploy them in environments with limited resources such as
|
57 |
+
a laptop, desktop or your own cloud infrastructure, democratizing access to
|
58 |
+
state of the art AI models and helping foster innovation for everyone.
|
59 |
+
|
60 |
+
### Usage
|
61 |
+
|
62 |
+
Below we share some code snippets on how to get quickly started with running the model. First make sure to `pip install -U transformers`, then copy the snippet from the section that is relevant for your usecase.
|
63 |
+
|
64 |
+
#### Running the model on a CPU
|
65 |
+
|
66 |
+
```python
|
67 |
+
from transformers import AutoTokenizer, AutoModelForCausalLM
|
68 |
+
|
69 |
+
tokenizer = AutoTokenizer.from_pretrained("beomi/gemma-ko-2b")
|
70 |
+
model = AutoModelForCausalLM.from_pretrained("beomi/gemma-ko-2b")
|
71 |
+
|
72 |
+
input_text = "머신러닝과 딥러닝의 차이는"
|
73 |
+
input_ids = tokenizer(input_text, return_tensors="pt")
|
74 |
+
|
75 |
+
outputs = model.generate(**input_ids)
|
76 |
+
print(tokenizer.decode(outputs[0]))
|
77 |
+
```
|
78 |
+
|
79 |
+
|
80 |
+
#### Running the model on a single / multi GPU
|
81 |
+
|
82 |
+
```python
|
83 |
+
# pip install accelerate
|
84 |
+
from transformers import AutoTokenizer, AutoModelForCausalLM
|
85 |
+
|
86 |
+
tokenizer = AutoTokenizer.from_pretrained("beomi/gemma-ko-2b")
|
87 |
+
model = AutoModelForCausalLM.from_pretrained("beomi/gemma-ko-2b", device_map="auto")
|
88 |
+
|
89 |
+
input_text = "머신러닝과 딥러닝의 차이는"
|
90 |
+
input_ids = tokenizer(input_text, return_tensors="pt").to("cuda")
|
91 |
+
|
92 |
+
outputs = model.generate(**input_ids)
|
93 |
+
print(tokenizer.decode(outputs[0]))
|
94 |
+
```
|
95 |
+
|
96 |
+
#### Other optimizations
|
97 |
+
|
98 |
+
* _Flash Attention 2_
|
99 |
+
|
100 |
+
First make sure to install `flash-attn` in your environment `pip install flash-attn`
|
101 |
+
|
102 |
+
```diff
|
103 |
+
model = AutoModelForCausalLM.from_pretrained(
|
104 |
+
"beomi/gemma-ko-2b",
|
105 |
+
torch_dtype=torch.float16,
|
106 |
+
+ attn_implementation="flash_attention_2"
|
107 |
+
).to(0)
|
108 |
+
```
|
109 |
+
|
110 |
+
### Inputs and outputs
|
111 |
+
|
112 |
+
* **Input:** Text string, such as a question, a prompt, or a document to be
|
113 |
+
summarized.
|
114 |
+
* **Output:** Generated Korean/English-language text in response to the input, such
|
115 |
+
as an answer to a question, or a summary of a document.
|
116 |
+
|
117 |
+
## Implementation Information
|
118 |
+
|
119 |
+
Details about the model internals.
|
120 |
+
|
121 |
+
### Software
|
122 |
+
|
123 |
+
Training was done using [beomi/Gemma-EasyLM](https://github.com/Beomi/Gemma-EasyLM).
|
124 |
+
|
125 |
+
|
126 |
+
## Evaluation
|
127 |
+
|
128 |
+
Model evaluation metrics and results.
|
129 |
+
|
130 |
+
### Benchmark Results
|
131 |
+
|
132 |
+
TBD
|
133 |
+
|
134 |
+
## Usage and Limitations
|
135 |
+
|
136 |
+
These models have certain limitations that users should be aware of.
|
137 |
+
|
138 |
+
### Intended Usage
|
139 |
+
|
140 |
+
Open Large Language Models (LLMs) have a wide range of applications across
|
141 |
+
various industries and domains. The following list of potential uses is not
|
142 |
+
comprehensive. The purpose of this list is to provide contextual information
|
143 |
+
about the possible use-cases that the model creators considered as part of model
|
144 |
+
training and development.
|
145 |
+
|
146 |
+
* Content Creation and Communication
|
147 |
+
* Text Generation: These models can be used to generate creative text formats
|
148 |
+
such as poems, scripts, code, marketing copy, and email drafts.
|
149 |
+
* Research and Education
|
150 |
+
* Natural Language Processing (NLP) Research: These models can serve as a
|
151 |
+
foundation for researchers to experiment with NLP techniques, develop
|
152 |
+
algorithms, and contribute to the advancement of the field.
|
153 |
+
* Language Learning Tools: Support interactive language learning experiences,
|
154 |
+
aiding in grammar correction or providing writing practice.
|
155 |
+
* Knowledge Exploration: Assist researchers in exploring large bodies of text
|
156 |
+
by generating summaries or answering questions about specific topics.
|
157 |
+
|
158 |
+
### Limitations
|
159 |
+
|
160 |
+
* Training Data
|
161 |
+
* The quality and diversity of the training data significantly influence the
|
162 |
+
model's capabilities. Biases or gaps in the training data can lead to
|
163 |
+
limitations in the model's responses.
|
164 |
+
* The scope of the training dataset determines the subject areas the model can
|
165 |
+
handle effectively.
|
166 |
+
* Context and Task Complexity
|
167 |
+
* LLMs are better at tasks that can be framed with clear prompts and
|
168 |
+
instructions. Open-ended or highly complex tasks might be challenging.
|
169 |
+
* A model's performance can be influenced by the amount of context provided
|
170 |
+
(longer context generally leads to better outputs, up to a certain point).
|
171 |
+
* Language Ambiguity and Nuance
|
172 |
+
* Natural language is inherently complex. LLMs might struggle to grasp subtle
|
173 |
+
nuances, sarcasm, or figurative language.
|
174 |
+
* Factual Accuracy
|
175 |
+
* LLMs generate responses based on information they learned from their
|
176 |
+
training datasets, but they are not knowledge bases. They may generate
|
177 |
+
incorrect or outdated factual statements.
|
178 |
+
* Common Sense
|
179 |
+
* LLMs rely on statistical patterns in language. They might lack the ability
|
180 |
+
to apply common sense reasoning in certain situations.
|
181 |
+
|
182 |
+
### Ethical Considerations and Risks
|
183 |
+
|
184 |
+
The development of large language models (LLMs) raises several ethical concerns.
|
185 |
+
In creating an open model, we have carefully considered the following:
|
186 |
+
|
187 |
+
* Bias and Fairness
|
188 |
+
* LLMs trained on large-scale, real-world text data can reflect socio-cultural
|
189 |
+
biases embedded in the training material. These models underwent careful
|
190 |
+
scrutiny, input data pre-processing described and posterior evaluations
|
191 |
+
reported in this card.
|
192 |
+
* Misinformation and Misuse
|
193 |
+
* LLMs can be misused to generate text that is false, misleading, or harmful.
|
194 |
+
* Guidelines are provided for responsible use with the model, see the
|
195 |
+
[Responsible Generative AI Toolkit](http://ai.google.dev/gemma/responsible).
|
196 |
+
* Transparency and Accountability:
|
197 |
+
* This model card summarizes details on the models' architecture,
|
198 |
+
capabilities, limitations, and evaluation processes.
|
199 |
+
* A responsibly developed open model offers the opportunity to share
|
200 |
+
innovation by making LLM technology accessible to developers and researchers
|
201 |
+
across the AI ecosystem.
|
202 |
+
|
203 |
+
Risks identified and mitigations:
|
204 |
+
|
205 |
+
* Perpetuation of biases: It's encouraged to perform continuous monitoring
|
206 |
+
(using evaluation metrics, human review) and the exploration of de-biasing
|
207 |
+
techniques during model training, fine-tuning, and other use cases.
|
208 |
+
* Generation of harmful content: Mechanisms and guidelines for content safety
|
209 |
+
are essential. Developers are encouraged to exercise caution and implement
|
210 |
+
appropriate content safety safeguards based on their specific product policies
|
211 |
+
and application use cases.
|
212 |
+
* Misuse for malicious purposes: Technical limitations and developer and
|
213 |
+
end-user education can help mitigate against malicious applications of LLMs.
|
214 |
+
Educational resources and reporting mechanisms for users to flag misuse are
|
215 |
+
provided. Prohibited uses of Gemma models are outlined in the
|
216 |
+
[Gemma Prohibited Use Policy](https://ai.google.dev/gemma/prohibited_use_policy).
|
217 |
+
* Privacy violations: Models were trained on data filtered for removal of PII
|
218 |
+
(Personally Identifiable Information). Developers are encouraged to adhere to
|
219 |
+
privacy regulations with privacy-preserving techniques.
|