Commit
•
a5d2cf7
1
Parent(s):
edbd328
uploaded readme
Browse files
README.md
ADDED
@@ -0,0 +1,320 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
Quantization made by Richard Erkhov.
|
2 |
+
|
3 |
+
[Github](https://github.com/RichardErkhov)
|
4 |
+
|
5 |
+
[Discord](https://discord.gg/pvy7H8DZMG)
|
6 |
+
|
7 |
+
[Request more models](https://github.com/RichardErkhov/quant_request)
|
8 |
+
|
9 |
+
|
10 |
+
gemma-mling-7b - GGUF
|
11 |
+
- Model creator: https://huggingface.co/beomi/
|
12 |
+
- Original model: https://huggingface.co/beomi/gemma-mling-7b/
|
13 |
+
|
14 |
+
|
15 |
+
| Name | Quant method | Size |
|
16 |
+
| ---- | ---- | ---- |
|
17 |
+
| [gemma-mling-7b.Q2_K.gguf](https://huggingface.co/RichardErkhov/beomi_-_gemma-mling-7b-gguf/blob/main/gemma-mling-7b.Q2_K.gguf) | Q2_K | 3.24GB |
|
18 |
+
| [gemma-mling-7b.IQ3_XS.gguf](https://huggingface.co/RichardErkhov/beomi_-_gemma-mling-7b-gguf/blob/main/gemma-mling-7b.IQ3_XS.gguf) | IQ3_XS | 3.54GB |
|
19 |
+
| [gemma-mling-7b.IQ3_S.gguf](https://huggingface.co/RichardErkhov/beomi_-_gemma-mling-7b-gguf/blob/main/gemma-mling-7b.IQ3_S.gguf) | IQ3_S | 3.71GB |
|
20 |
+
| [gemma-mling-7b.Q3_K_S.gguf](https://huggingface.co/RichardErkhov/beomi_-_gemma-mling-7b-gguf/blob/main/gemma-mling-7b.Q3_K_S.gguf) | Q3_K_S | 3.71GB |
|
21 |
+
| [gemma-mling-7b.IQ3_M.gguf](https://huggingface.co/RichardErkhov/beomi_-_gemma-mling-7b-gguf/blob/main/gemma-mling-7b.IQ3_M.gguf) | IQ3_M | 3.82GB |
|
22 |
+
| [gemma-mling-7b.Q3_K.gguf](https://huggingface.co/RichardErkhov/beomi_-_gemma-mling-7b-gguf/blob/main/gemma-mling-7b.Q3_K.gguf) | Q3_K | 4.07GB |
|
23 |
+
| [gemma-mling-7b.Q3_K_M.gguf](https://huggingface.co/RichardErkhov/beomi_-_gemma-mling-7b-gguf/blob/main/gemma-mling-7b.Q3_K_M.gguf) | Q3_K_M | 4.07GB |
|
24 |
+
| [gemma-mling-7b.Q3_K_L.gguf](https://huggingface.co/RichardErkhov/beomi_-_gemma-mling-7b-gguf/blob/main/gemma-mling-7b.Q3_K_L.gguf) | Q3_K_L | 4.39GB |
|
25 |
+
| [gemma-mling-7b.IQ4_XS.gguf](https://huggingface.co/RichardErkhov/beomi_-_gemma-mling-7b-gguf/blob/main/gemma-mling-7b.IQ4_XS.gguf) | IQ4_XS | 4.48GB |
|
26 |
+
| [gemma-mling-7b.Q4_0.gguf](https://huggingface.co/RichardErkhov/beomi_-_gemma-mling-7b-gguf/blob/main/gemma-mling-7b.Q4_0.gguf) | Q4_0 | 4.67GB |
|
27 |
+
| [gemma-mling-7b.IQ4_NL.gguf](https://huggingface.co/RichardErkhov/beomi_-_gemma-mling-7b-gguf/blob/main/gemma-mling-7b.IQ4_NL.gguf) | IQ4_NL | 4.69GB |
|
28 |
+
| [gemma-mling-7b.Q4_K_S.gguf](https://huggingface.co/RichardErkhov/beomi_-_gemma-mling-7b-gguf/blob/main/gemma-mling-7b.Q4_K_S.gguf) | Q4_K_S | 4.7GB |
|
29 |
+
| [gemma-mling-7b.Q4_K.gguf](https://huggingface.co/RichardErkhov/beomi_-_gemma-mling-7b-gguf/blob/main/gemma-mling-7b.Q4_K.gguf) | Q4_K | 4.96GB |
|
30 |
+
| [gemma-mling-7b.Q4_K_M.gguf](https://huggingface.co/RichardErkhov/beomi_-_gemma-mling-7b-gguf/blob/main/gemma-mling-7b.Q4_K_M.gguf) | Q4_K_M | 4.96GB |
|
31 |
+
| [gemma-mling-7b.Q4_1.gguf](https://huggingface.co/RichardErkhov/beomi_-_gemma-mling-7b-gguf/blob/main/gemma-mling-7b.Q4_1.gguf) | Q4_1 | 5.12GB |
|
32 |
+
| [gemma-mling-7b.Q5_0.gguf](https://huggingface.co/RichardErkhov/beomi_-_gemma-mling-7b-gguf/blob/main/gemma-mling-7b.Q5_0.gguf) | Q5_0 | 5.57GB |
|
33 |
+
| [gemma-mling-7b.Q5_K_S.gguf](https://huggingface.co/RichardErkhov/beomi_-_gemma-mling-7b-gguf/blob/main/gemma-mling-7b.Q5_K_S.gguf) | Q5_K_S | 5.57GB |
|
34 |
+
| [gemma-mling-7b.Q5_K.gguf](https://huggingface.co/RichardErkhov/beomi_-_gemma-mling-7b-gguf/blob/main/gemma-mling-7b.Q5_K.gguf) | Q5_K | 5.72GB |
|
35 |
+
| [gemma-mling-7b.Q5_K_M.gguf](https://huggingface.co/RichardErkhov/beomi_-_gemma-mling-7b-gguf/blob/main/gemma-mling-7b.Q5_K_M.gguf) | Q5_K_M | 5.72GB |
|
36 |
+
| [gemma-mling-7b.Q5_1.gguf](https://huggingface.co/RichardErkhov/beomi_-_gemma-mling-7b-gguf/blob/main/gemma-mling-7b.Q5_1.gguf) | Q5_1 | 6.02GB |
|
37 |
+
| [gemma-mling-7b.Q6_K.gguf](https://huggingface.co/RichardErkhov/beomi_-_gemma-mling-7b-gguf/blob/main/gemma-mling-7b.Q6_K.gguf) | Q6_K | 6.53GB |
|
38 |
+
| [gemma-mling-7b.Q8_0.gguf](https://huggingface.co/RichardErkhov/beomi_-_gemma-mling-7b-gguf/blob/main/gemma-mling-7b.Q8_0.gguf) | Q8_0 | 8.45GB |
|
39 |
+
|
40 |
+
|
41 |
+
|
42 |
+
|
43 |
+
Original model description:
|
44 |
+
---
|
45 |
+
language:
|
46 |
+
- ko
|
47 |
+
- en
|
48 |
+
- zh
|
49 |
+
- ja
|
50 |
+
license: other
|
51 |
+
library_name: transformers
|
52 |
+
license_name: gemma-terms-of-use
|
53 |
+
license_link: https://ai.google.dev/gemma/terms
|
54 |
+
pipeline_tag: text-generation
|
55 |
+
tags:
|
56 |
+
- pytorch
|
57 |
+
---
|
58 |
+
|
59 |
+
# Gemma-Mling: Multilingual Gemma
|
60 |
+
|
61 |
+
> Update @ 2024.04.15: First release of Gemma-Mling 7B model
|
62 |
+
|
63 |
+
**Original Gemma Model Page**: [Gemma](https://ai.google.dev/gemma/docs)
|
64 |
+
|
65 |
+
This model card corresponds to the 7B base version of the **Gemma-Mling** model,
|
66 |
+
continual pretrained on mainly Korean/English/Chinese/Japanese + 500 multilingual corpus.
|
67 |
+
|
68 |
+
**Resources and Technical Documentation**:
|
69 |
+
|
70 |
+
* [Original Google's Gemma-7B](https://huggingface.co/google/gemma-7b)
|
71 |
+
* [Training Code @ Github: Gemma-EasyLM](https://github.com/Beomi/Gemma-EasyLM)
|
72 |
+
|
73 |
+
**Terms of Use**: [Terms](https://www.kaggle.com/models/google/gemma/license/consent)
|
74 |
+
|
75 |
+
**Citation**
|
76 |
+
|
77 |
+
```bibtex
|
78 |
+
@misc {gemma_mling_7b,
|
79 |
+
author = { {Junbum Lee, Taekyoon Choi} },
|
80 |
+
title = { gemma-mling-7b },
|
81 |
+
year = 2024,
|
82 |
+
url = { https://huggingface.co/beomi/gemma-mling-7b },
|
83 |
+
publisher = { Hugging Face }
|
84 |
+
}
|
85 |
+
```
|
86 |
+
|
87 |
+
**Model Developers**: Junbum Lee (Beomi) & Taekyoon Choi (Taekyoon)
|
88 |
+
|
89 |
+
## Model Information
|
90 |
+
|
91 |
+
### Usage
|
92 |
+
|
93 |
+
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.
|
94 |
+
|
95 |
+
#### Running the model on a CPU
|
96 |
+
|
97 |
+
```python
|
98 |
+
from transformers import AutoTokenizer, AutoModelForCausalLM
|
99 |
+
|
100 |
+
tokenizer = AutoTokenizer.from_pretrained("beomi/gemma-mling-7b")
|
101 |
+
model = AutoModelForCausalLM.from_pretrained("beomi/gemma-mling-7b")
|
102 |
+
|
103 |
+
input_text = "머신러닝과 딥러닝의 차이는"
|
104 |
+
input_ids = tokenizer(input_text, return_tensors="pt")
|
105 |
+
|
106 |
+
outputs = model.generate(**input_ids)
|
107 |
+
print(tokenizer.decode(outputs[0]))
|
108 |
+
```
|
109 |
+
|
110 |
+
|
111 |
+
#### Running the model on a single / multi GPU
|
112 |
+
|
113 |
+
```python
|
114 |
+
# pip install accelerate
|
115 |
+
from transformers import AutoTokenizer, AutoModelForCausalLM
|
116 |
+
|
117 |
+
tokenizer = AutoTokenizer.from_pretrained("beomi/gemma-mling-7b")
|
118 |
+
model = AutoModelForCausalLM.from_pretrained("beomi/gemma-mling-7b", device_map="auto")
|
119 |
+
|
120 |
+
input_text = "머신러닝과 딥러닝의 차이는"
|
121 |
+
input_ids = tokenizer(input_text, return_tensors="pt").to("cuda")
|
122 |
+
|
123 |
+
outputs = model.generate(**input_ids)
|
124 |
+
print(tokenizer.decode(outputs[0]))
|
125 |
+
```
|
126 |
+
|
127 |
+
### Inputs and outputs
|
128 |
+
|
129 |
+
* **Input:** Text string, such as a question, a prompt, or a document to be
|
130 |
+
summarized.
|
131 |
+
* **Output:** Generated Multilingual-language text in response to the input, such
|
132 |
+
as an answer to a question, or a summary of a document.
|
133 |
+
|
134 |
+
## Implementation Information
|
135 |
+
|
136 |
+
Details about the model internals.
|
137 |
+
|
138 |
+
### Software
|
139 |
+
|
140 |
+
Training was done using [beomi/Gemma-EasyLM](https://github.com/Beomi/Gemma-EasyLM).
|
141 |
+
|
142 |
+
### Dataset
|
143 |
+
|
144 |
+
We trained a mixture of multiple language datasets and trained until 100B.
|
145 |
+
The released model is the best performance model based on our Evaluation below from model checkpoints.
|
146 |
+
|
147 |
+
For Korean and English datasets, we utilized sampled llama2ko training dataset which combined 1:1 ratio in each language.
|
148 |
+
|
149 |
+
| Dataset | Jsonl (GB) | Sampled |
|
150 |
+
|--------------------------|------------|---------|
|
151 |
+
| range3/cc100-ja | 96.39 | No |
|
152 |
+
| Skywork/SkyPile-150B | 100.57 | Yes |
|
153 |
+
| llama2ko dataset (ko/en) | 108.5 | Yes |
|
154 |
+
| cis-lmu/Glot500 | 181.24 | No |
|
155 |
+
| Total | 486.7 | . |
|
156 |
+
|
157 |
+
## Training Progress
|
158 |
+
|
159 |
+
- Report Link: https://api.wandb.ai/links/tgchoi/6lt0ce3s
|
160 |
+
|
161 |
+
## Evaluation
|
162 |
+
|
163 |
+
Model evaluation metrics and results.
|
164 |
+
|
165 |
+
### Evaluation Scripts
|
166 |
+
|
167 |
+
- For Knowledge / KoBest / XCOPA / XWinograd
|
168 |
+
- [EleutherAI/lm-evaluation-harness](https://github.com/EleutherAI/lm-evaluation-harness) v0.4.2
|
169 |
+
```bash
|
170 |
+
!git clone https://github.com/EleutherAI/lm-evaluation-harness.git
|
171 |
+
!cd lm-evaluation-harness && pip install -r requirements.txt && pip install -e .
|
172 |
+
|
173 |
+
!lm_eval --model hf \
|
174 |
+
--model_args pretrained=beomi/gemma-mling-7b,dtype="float16" \
|
175 |
+
--tasks "haerae,kobest,kmmlu_direct,cmmlu,ceval-valid,mmlu,xwinograd,xcopa \
|
176 |
+
--num_fewshot "0,5,5,5,5,5,0,5" \
|
177 |
+
--device cuda
|
178 |
+
```
|
179 |
+
- For JP Eval Harness
|
180 |
+
- [Stability-AI/lm-evaluation-harness (`jp-stable` branch)](https://github.com/Stability-AI/lm-evaluation-harness/tree/jp-stable)
|
181 |
+
```bash
|
182 |
+
!git clone -b jp-stable https://github.com/Stability-AI/lm-evaluation-harness.git
|
183 |
+
!cd lm-evaluation-harness && pip install -e ".[ja]"
|
184 |
+
!pip install 'fugashi[unidic]' && python -m unidic download
|
185 |
+
|
186 |
+
!cd lm-evaluation-harness && python main.py \
|
187 |
+
--model hf-causal \
|
188 |
+
--model_args pretrained=beomi/gemma-mling-7b,torch_dtype='auto'"
|
189 |
+
--tasks "jcommonsenseqa-1.1-0.3,jnli-1.3-0.3,marc_ja-1.1-0.3,jsquad-1.1-0.3,jaqket_v2-0.2-0.3,xlsum_ja,mgsm"
|
190 |
+
--num_fewshot "3,3,3,2,1,1,5"
|
191 |
+
```
|
192 |
+
|
193 |
+
### Benchmark Results
|
194 |
+
|
195 |
+
| Category | Metric | Shots | Score |
|
196 |
+
|----------------------------------|----------------------|------------|--------|
|
197 |
+
| **Default Metric** | **ACC** | | |
|
198 |
+
| **Knowledge (5-shot)** | MMLU | | 61.76 |
|
199 |
+
| | KMMLU (Exact Match) | | 42.75 |
|
200 |
+
| | CMLU | | 50.93 |
|
201 |
+
| | JMLU | | |
|
202 |
+
| | C-EVAL | | 50.07 |
|
203 |
+
| | HAERAE | 0-shot | 63.89 |
|
204 |
+
| **KoBest (5-shot)** | BoolQ | | 85.47 |
|
205 |
+
| | COPA | | 83.5 |
|
206 |
+
| | Hellaswag (acc-norm) | | 63.2 |
|
207 |
+
| | Sentineg | | 97.98 |
|
208 |
+
| | WiC | | 70.95 |
|
209 |
+
| **XCOPA (5-shot)** | IT | | 72.8 |
|
210 |
+
| | ID | | 76.4 |
|
211 |
+
| | TH | | 60.2 |
|
212 |
+
| | TR | | 65.6 |
|
213 |
+
| | VI | | 77.2 |
|
214 |
+
| | ZH | | 80.2 |
|
215 |
+
| **JP Eval Harness (Prompt ver 0.3)** | JcommonsenseQA | 3-shot | 85.97 |
|
216 |
+
| | JNLI | 3-shot | 39.11 |
|
217 |
+
| | Marc_ja | 3-shot | 96.48 |
|
218 |
+
| | JSquad (Exact Match) | 2-shot | 70.69 |
|
219 |
+
| | Jaqket (Exact Match) | 1-shot | 81.53 |
|
220 |
+
| | MGSM | 5-shot | 28.8 |
|
221 |
+
| **XWinograd (0-shot)** | EN | | 89.03 |
|
222 |
+
| | FR | | 72.29 |
|
223 |
+
| | JP | | 82.69 |
|
224 |
+
| | PT | | 73.38 |
|
225 |
+
| | RU | | 68.57 |
|
226 |
+
| | ZH | | 79.17 |
|
227 |
+
|
228 |
+
|
229 |
+
|
230 |
+
## Usage and Limitations
|
231 |
+
|
232 |
+
These models have certain limitations that users should be aware of.
|
233 |
+
|
234 |
+
### Intended Usage
|
235 |
+
|
236 |
+
Open Large Language Models (LLMs) have a wide range of applications across
|
237 |
+
various industries and domains. The following list of potential uses is not
|
238 |
+
comprehensive. The purpose of this list is to provide contextual information
|
239 |
+
about the possible use-cases that the model creators considered as part of model
|
240 |
+
training and development.
|
241 |
+
|
242 |
+
* Content Creation and Communication
|
243 |
+
* Text Generation: These models can be used to generate creative text formats
|
244 |
+
such as poems, scripts, code, marketing copy, and email drafts.
|
245 |
+
* Research and Education
|
246 |
+
* Natural Language Processing (NLP) Research: These models can serve as a
|
247 |
+
foundation for researchers to experiment with NLP techniques, develop
|
248 |
+
algorithms, and contribute to the advancement of the field.
|
249 |
+
* Language Learning Tools: Support interactive language learning experiences,
|
250 |
+
aiding in grammar correction or providing writing practice.
|
251 |
+
* Knowledge Exploration: Assist researchers in exploring large bodies of text
|
252 |
+
by generating summaries or answering questions about specific topics.
|
253 |
+
|
254 |
+
### Limitations
|
255 |
+
|
256 |
+
* Training Data
|
257 |
+
* The quality and diversity of the training data significantly influence the
|
258 |
+
model's capabilities. Biases or gaps in the training data can lead to
|
259 |
+
limitations in the model's responses.
|
260 |
+
* The scope of the training dataset determines the subject areas the model can
|
261 |
+
handle effectively.
|
262 |
+
* Context and Task Complexity
|
263 |
+
* LLMs are better at tasks that can be framed with clear prompts and
|
264 |
+
instructions. Open-ended or highly complex tasks might be challenging.
|
265 |
+
* A model's performance can be influenced by the amount of context provided
|
266 |
+
(longer context generally leads to better outputs, up to a certain point).
|
267 |
+
* Language Ambiguity and Nuance
|
268 |
+
* Natural language is inherently complex. LLMs might struggle to grasp subtle
|
269 |
+
nuances, sarcasm, or figurative language.
|
270 |
+
* Factual Accuracy
|
271 |
+
* LLMs generate responses based on information they learned from their
|
272 |
+
training datasets, but they are not knowledge bases. They may generate
|
273 |
+
incorrect or outdated factual statements.
|
274 |
+
* Common Sense
|
275 |
+
* LLMs rely on statistical patterns in language. They might lack the ability
|
276 |
+
to apply common sense reasoning in certain situations.
|
277 |
+
|
278 |
+
### Ethical Considerations and Risks
|
279 |
+
|
280 |
+
The development of large language models (LLMs) raises several ethical concerns.
|
281 |
+
In creating an open model, we have carefully considered the following:
|
282 |
+
|
283 |
+
* Bias and Fairness
|
284 |
+
* LLMs trained on large-scale, real-world text data can reflect socio-cultural
|
285 |
+
biases embedded in the training material. These models underwent careful
|
286 |
+
scrutiny, input data pre-processing described and posterior evaluations
|
287 |
+
reported in this card.
|
288 |
+
* Misinformation and Misuse
|
289 |
+
* LLMs can be misused to generate text that is false, misleading, or harmful.
|
290 |
+
* Guidelines are provided for responsible use with the model, see the
|
291 |
+
[Responsible Generative AI Toolkit](http://ai.google.dev/gemma/responsible).
|
292 |
+
* Transparency and Accountability:
|
293 |
+
* This model card summarizes details on the models' architecture,
|
294 |
+
capabilities, limitations, and evaluation processes.
|
295 |
+
* A responsibly developed open model offers the opportunity to share
|
296 |
+
innovation by making LLM technology accessible to developers and researchers
|
297 |
+
across the AI ecosystem.
|
298 |
+
|
299 |
+
Risks identified and mitigations:
|
300 |
+
|
301 |
+
* Perpetuation of biases: It's encouraged to perform continuous monitoring
|
302 |
+
(using evaluation metrics, human review) and the exploration of de-biasing
|
303 |
+
techniques during model training, fine-tuning, and other use cases.
|
304 |
+
* Generation of harmful content: Mechanisms and guidelines for content safety
|
305 |
+
are essential. Developers are encouraged to exercise caution and implement
|
306 |
+
appropriate content safety safeguards based on their specific product policies
|
307 |
+
and application use cases.
|
308 |
+
* Misuse for malicious purposes: Technical limitations and developer and
|
309 |
+
end-user education can help mitigate against malicious applications of LLMs.
|
310 |
+
Educational resources and reporting mechanisms for users to flag misuse are
|
311 |
+
provided. Prohibited uses of Gemma models are outlined in the
|
312 |
+
[Gemma Prohibited Use Policy](https://ai.google.dev/gemma/prohibited_use_policy).
|
313 |
+
* Privacy violations: Models were trained on data filtered for removal of PII
|
314 |
+
(Personally Identifiable Information). Developers are encouraged to adhere to
|
315 |
+
privacy regulations with privacy-preserving techniques.
|
316 |
+
|
317 |
+
## Acknowledgement
|
318 |
+
|
319 |
+
The training is supported by [TPU Research Cloud](https://sites.research.google/trc/) program.
|
320 |
+
|