Update README.md
Browse files
README.md
CHANGED
@@ -1,199 +1,222 @@
|
|
1 |
---
|
|
|
|
|
|
|
|
|
2 |
library_name: transformers
|
3 |
-
|
|
|
|
|
|
|
|
|
4 |
---
|
5 |
|
6 |
-
#
|
7 |
|
8 |
-
|
|
|
9 |
|
|
|
10 |
|
|
|
11 |
|
12 |
-
|
|
|
13 |
|
14 |
-
|
15 |
|
16 |
-
|
17 |
|
18 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
19 |
|
20 |
-
|
21 |
-
- **Funded by [optional]:** [More Information Needed]
|
22 |
-
- **Shared by [optional]:** [More Information Needed]
|
23 |
-
- **Model type:** [More Information Needed]
|
24 |
-
- **Language(s) (NLP):** [More Information Needed]
|
25 |
-
- **License:** [More Information Needed]
|
26 |
-
- **Finetuned from model [optional]:** [More Information Needed]
|
27 |
|
28 |
-
|
29 |
|
30 |
-
|
31 |
|
32 |
-
|
33 |
-
- **Paper [optional]:** [More Information Needed]
|
34 |
-
- **Demo [optional]:** [More Information Needed]
|
35 |
|
36 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
37 |
|
38 |
-
|
39 |
|
40 |
-
|
41 |
|
42 |
-
|
43 |
|
44 |
-
|
|
|
45 |
|
46 |
-
|
|
|
47 |
|
48 |
-
|
|
|
49 |
|
50 |
-
|
|
|
|
|
51 |
|
52 |
-
### Out-of-Scope Use
|
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 |
-
### Training Procedure
|
85 |
-
|
86 |
-
<!-- This relates heavily to the Technical Specifications. Content here should link to that section when it is relevant to the training procedure. -->
|
87 |
-
|
88 |
-
#### Preprocessing [optional]
|
89 |
-
|
90 |
-
[More Information Needed]
|
91 |
-
|
92 |
-
|
93 |
-
#### Training Hyperparameters
|
94 |
-
|
95 |
-
- **Training regime:** [More Information Needed] <!--fp32, fp16 mixed precision, bf16 mixed precision, bf16 non-mixed precision, fp16 non-mixed precision, fp8 mixed precision -->
|
96 |
-
|
97 |
-
#### Speeds, Sizes, Times [optional]
|
98 |
-
|
99 |
-
<!-- This section provides information about throughput, start/end time, checkpoint size if relevant, etc. -->
|
100 |
-
|
101 |
-
[More Information Needed]
|
102 |
|
103 |
## Evaluation
|
104 |
|
105 |
-
|
106 |
-
|
107 |
-
###
|
108 |
-
|
109 |
-
|
110 |
-
|
111 |
-
|
112 |
-
|
113 |
-
|
114 |
-
|
115 |
-
|
116 |
-
|
117 |
-
|
118 |
-
|
119 |
-
|
120 |
-
|
121 |
-
|
122 |
-
|
123 |
-
|
124 |
-
|
125 |
-
|
126 |
-
|
127 |
-
|
128 |
-
|
129 |
-
|
130 |
-
|
131 |
-
|
132 |
-
|
133 |
-
|
134 |
-
|
135 |
-
|
136 |
-
|
137 |
-
|
138 |
-
|
139 |
-
|
140 |
-
|
141 |
-
|
142 |
-
|
143 |
-
|
144 |
-
|
145 |
-
|
146 |
-
|
147 |
-
|
148 |
-
|
149 |
-
|
150 |
-
|
151 |
-
|
152 |
-
|
153 |
-
|
154 |
-
|
155 |
-
|
156 |
-
|
157 |
-
|
158 |
-
|
159 |
-
###
|
160 |
-
|
161 |
-
|
162 |
-
|
163 |
-
|
164 |
-
|
165 |
-
|
166 |
-
|
167 |
-
|
168 |
-
|
169 |
-
|
170 |
-
|
171 |
-
|
172 |
-
|
173 |
-
|
174 |
-
|
175 |
-
|
176 |
-
|
177 |
-
|
178 |
-
|
179 |
-
|
180 |
-
|
181 |
-
|
182 |
-
|
183 |
-
|
184 |
-
|
185 |
-
|
186 |
-
|
187 |
-
|
188 |
-
|
189 |
-
|
190 |
-
|
191 |
-
|
192 |
-
|
193 |
-
|
194 |
-
|
195 |
-
|
196 |
-
|
197 |
-
|
198 |
-
|
199 |
-
|
|
|
|
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 |
+
**Quantuzed from:**: [Gemma Ko](https://huggingface.co/beomi/gemma-ko-7b)
|
17 |
+
**Original Gemma Model Page**: [Gemma](https://ai.google.dev/gemma/docs)
|
18 |
|
19 |
+
This model card corresponds to the 7B base version of the **Gemma-Ko** model.
|
20 |
|
21 |
+
**Resources and Technical Documentation**:
|
22 |
|
23 |
+
* [Original Google's Gemma-7B](https://huggingface.co/google/gemma-7b)
|
24 |
+
* [Training Code @ Github: Gemma-EasyLM](https://github.com/Beomi/Gemma-EasyLM)
|
25 |
|
26 |
+
**Terms of Use**: [Terms](https://www.kaggle.com/models/google/gemma/license/consent)
|
27 |
|
28 |
+
**Citation**
|
29 |
|
30 |
+
```bibtex
|
31 |
+
@misc {gemma_ko_7b,
|
32 |
+
author = { {Junbum Lee, Taekyoon Choi} },
|
33 |
+
title = { gemma-ko-7b },
|
34 |
+
year = 2024,
|
35 |
+
url = { https://huggingface.co/beomi/gemma-ko-7b },
|
36 |
+
doi = { 10.57967/hf/1859 },
|
37 |
+
publisher = { Hugging Face }
|
38 |
+
}
|
39 |
+
```
|
40 |
|
41 |
+
**Model Developers**: Junbum Lee (Beomi) & Taekyoon Choi (Taekyoon)
|
|
|
|
|
|
|
|
|
|
|
|
|
42 |
|
43 |
+
## Model Information
|
44 |
|
45 |
+
Summary description and brief definition of inputs and outputs.
|
46 |
|
47 |
+
### Description
|
|
|
|
|
48 |
|
49 |
+
Gemma is a family of lightweight, state-of-the-art open models from Google,
|
50 |
+
built from the same research and technology used to create the Gemini models.
|
51 |
+
They are text-to-text, decoder-only large language models, available in English,
|
52 |
+
with open weights, pre-trained variants, and instruction-tuned variants. Gemma
|
53 |
+
models are well-suited for a variety of text generation tasks, including
|
54 |
+
question answering, summarization, and reasoning. Their relatively small size
|
55 |
+
makes it possible to deploy them in environments with limited resources such as
|
56 |
+
a laptop, desktop or your own cloud infrastructure, democratizing access to
|
57 |
+
state of the art AI models and helping foster innovation for everyone.
|
58 |
|
59 |
+
### Usage
|
60 |
|
61 |
+
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.
|
62 |
|
63 |
+
#### Running the model on a CPU
|
64 |
|
65 |
+
```python
|
66 |
+
from transformers import AutoTokenizer, AutoModelForCausalLM
|
67 |
|
68 |
+
tokenizer = AutoTokenizer.from_pretrained("beomi/gemma-ko-7b")
|
69 |
+
model = AutoModelForCausalLM.from_pretrained("beomi/gemma-ko-7b")
|
70 |
|
71 |
+
input_text = "머신러닝과 딥러닝의 차이는"
|
72 |
+
input_ids = tokenizer(input_text, return_tensors="pt")
|
73 |
|
74 |
+
outputs = model.generate(**input_ids)
|
75 |
+
print(tokenizer.decode(outputs[0]))
|
76 |
+
```
|
77 |
|
|
|
78 |
|
79 |
+
#### Running the model on a single / multi GPU
|
80 |
|
81 |
+
```python
|
82 |
+
# pip install accelerate
|
83 |
+
from transformers import AutoTokenizer, AutoModelForCausalLM
|
84 |
|
85 |
+
tokenizer = AutoTokenizer.from_pretrained("beomi/gemma-ko-7b")
|
86 |
+
model = AutoModelForCausalLM.from_pretrained("beomi/gemma-ko-7b", device_map="auto")
|
87 |
|
88 |
+
input_text = "머신러닝과 딥러닝의 차이는"
|
89 |
+
input_ids = tokenizer(input_text, return_tensors="pt").to("cuda")
|
90 |
|
91 |
+
outputs = model.generate(**input_ids)
|
92 |
+
print(tokenizer.decode(outputs[0]))
|
93 |
+
```
|
94 |
|
95 |
+
#### Other optimizations
|
96 |
|
97 |
+
* _Flash Attention 2_
|
98 |
|
99 |
+
First make sure to install `flash-attn` in your environment `pip install flash-attn`
|
100 |
|
101 |
+
```diff
|
102 |
+
model = AutoModelForCausalLM.from_pretrained(
|
103 |
+
"beomi/gemma-ko-7b",
|
104 |
+
torch_dtype=torch.float16,
|
105 |
+
+ attn_implementation="flash_attention_2"
|
106 |
+
).to(0)
|
107 |
+
```
|
108 |
|
109 |
+
### Inputs and outputs
|
110 |
|
111 |
+
* **Input:** Text string, such as a question, a prompt, or a document to be
|
112 |
+
summarized.
|
113 |
+
* **Output:** Generated Korean/English-language text in response to the input, such
|
114 |
+
as an answer to a question, or a summary of a document.
|
115 |
|
116 |
+
## Implementation Information
|
117 |
|
118 |
+
Details about the model internals.
|
119 |
|
120 |
+
### Software
|
121 |
|
122 |
+
Training was done using [beomi/Gemma-EasyLM](https://github.com/Beomi/Gemma-EasyLM).
|
123 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
124 |
|
125 |
## Evaluation
|
126 |
|
127 |
+
Model evaluation metrics and results.
|
128 |
+
|
129 |
+
### Benchmark Results
|
130 |
+
|
131 |
+
TBD
|
132 |
+
|
133 |
+
## Usage and Limitations
|
134 |
+
|
135 |
+
These models have certain limitations that users should be aware of.
|
136 |
+
|
137 |
+
### Intended Usage
|
138 |
+
|
139 |
+
Open Large Language Models (LLMs) have a wide range of applications across
|
140 |
+
various industries and domains. The following list of potential uses is not
|
141 |
+
comprehensive. The purpose of this list is to provide contextual information
|
142 |
+
about the possible use-cases that the model creators considered as part of model
|
143 |
+
training and development.
|
144 |
+
|
145 |
+
* Content Creation and Communication
|
146 |
+
* Text Generation: These models can be used to generate creative text formats
|
147 |
+
such as poems, scripts, code, marketing copy, and email drafts.
|
148 |
+
* Research and Education
|
149 |
+
* Natural Language Processing (NLP) Research: These models can serve as a
|
150 |
+
foundation for researchers to experiment with NLP techniques, develop
|
151 |
+
algorithms, and contribute to the advancement of the field.
|
152 |
+
* Language Learning Tools: Support interactive language learning experiences,
|
153 |
+
aiding in grammar correction or providing writing practice.
|
154 |
+
* Knowledge Exploration: Assist researchers in exploring large bodies of text
|
155 |
+
by generating summaries or answering questions about specific topics.
|
156 |
+
|
157 |
+
### Limitations
|
158 |
+
|
159 |
+
* Training Data
|
160 |
+
* The quality and diversity of the training data significantly influence the
|
161 |
+
model's capabilities. Biases or gaps in the training data can lead to
|
162 |
+
limitations in the model's responses.
|
163 |
+
* The scope of the training dataset determines the subject areas the model can
|
164 |
+
handle effectively.
|
165 |
+
* Context and Task Complexity
|
166 |
+
* LLMs are better at tasks that can be framed with clear prompts and
|
167 |
+
instructions. Open-ended or highly complex tasks might be challenging.
|
168 |
+
* A model's performance can be influenced by the amount of context provided
|
169 |
+
(longer context generally leads to better outputs, up to a certain point).
|
170 |
+
* Language Ambiguity and Nuance
|
171 |
+
* Natural language is inherently complex. LLMs might struggle to grasp subtle
|
172 |
+
nuances, sarcasm, or figurative language.
|
173 |
+
* Factual Accuracy
|
174 |
+
* LLMs generate responses based on information they learned from their
|
175 |
+
training datasets, but they are not knowledge bases. They may generate
|
176 |
+
incorrect or outdated factual statements.
|
177 |
+
* Common Sense
|
178 |
+
* LLMs rely on statistical patterns in language. They might lack the ability
|
179 |
+
to apply common sense reasoning in certain situations.
|
180 |
+
|
181 |
+
### Ethical Considerations and Risks
|
182 |
+
|
183 |
+
The development of large language models (LLMs) raises several ethical concerns.
|
184 |
+
In creating an open model, we have carefully considered the following:
|
185 |
+
|
186 |
+
* Bias and Fairness
|
187 |
+
* LLMs trained on large-scale, real-world text data can reflect socio-cultural
|
188 |
+
biases embedded in the training material. These models underwent careful
|
189 |
+
scrutiny, input data pre-processing described and posterior evaluations
|
190 |
+
reported in this card.
|
191 |
+
* Misinformation and Misuse
|
192 |
+
* LLMs can be misused to generate text that is false, misleading, or harmful.
|
193 |
+
* Guidelines are provided for responsible use with the model, see the
|
194 |
+
[Responsible Generative AI Toolkit](http://ai.google.dev/gemma/responsible).
|
195 |
+
* Transparency and Accountability:
|
196 |
+
* This model card summarizes details on the models' architecture,
|
197 |
+
capabilities, limitations, and evaluation processes.
|
198 |
+
* A responsibly developed open model offers the opportunity to share
|
199 |
+
innovation by making LLM technology accessible to developers and researchers
|
200 |
+
across the AI ecosystem.
|
201 |
+
|
202 |
+
Risks identified and mitigations:
|
203 |
+
|
204 |
+
* Perpetuation of biases: It's encouraged to perform continuous monitoring
|
205 |
+
(using evaluation metrics, human review) and the exploration of de-biasing
|
206 |
+
techniques during model training, fine-tuning, and other use cases.
|
207 |
+
* Generation of harmful content: Mechanisms and guidelines for content safety
|
208 |
+
are essential. Developers are encouraged to exercise caution and implement
|
209 |
+
appropriate content safety safeguards based on their specific product policies
|
210 |
+
and application use cases.
|
211 |
+
* Misuse for malicious purposes: Technical limitations and developer and
|
212 |
+
end-user education can help mitigate against malicious applications of LLMs.
|
213 |
+
Educational resources and reporting mechanisms for users to flag misuse are
|
214 |
+
provided. Prohibited uses of Gemma models are outlined in the
|
215 |
+
[Gemma Prohibited Use Policy](https://ai.google.dev/gemma/prohibited_use_policy).
|
216 |
+
* Privacy violations: Models were trained on data filtered for removal of PII
|
217 |
+
(Personally Identifiable Information). Developers are encouraged to adhere to
|
218 |
+
privacy regulations with privacy-preserving techniques.
|
219 |
+
|
220 |
+
## Acknowledgement
|
221 |
+
|
222 |
+
The training is supported by [TPU Research Cloud](https://sites.research.google/trc/) program.
|