TechxGenus commited on
Commit
6fc27c8
1 Parent(s): 87a2c4c

Upload README.md

Browse files
Files changed (1) hide show
  1. README.md +450 -0
README.md ADDED
@@ -0,0 +1,450 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ library_name: transformers
3
+ tags: []
4
+ extra_gated_heading: "Access Gemma on Hugging Face"
5
+ extra_gated_prompt: "To access Gemma on Hugging Face, you’re required to review and agree to Google’s usage license. To do this, please ensure you’re logged-in to Hugging Face and click below. Requests are processed immediately."
6
+ extra_gated_button_content: "Acknowledge license"
7
+ license: other
8
+ license_name: gemma-terms-of-use
9
+ license_link: https://ai.google.dev/gemma/terms
10
+ ---
11
+
12
+ AWQ quantized version of gemma-7b model.
13
+
14
+ ---
15
+
16
+ # Gemma Model Card
17
+
18
+ **Model Page**: [Gemma](https://ai.google.dev/gemma/docs)
19
+
20
+ This model card corresponds to the 7B base version of the Gemma model. You can also visit the model card of the [2B base model](https://huggingface.co/google/gemma-2b), [7B instruct model](https://huggingface.co/google/gemma-7b-it), and [2B instruct model](https://huggingface.co/google/gemma-2b-it).
21
+
22
+ **Resources and Technical Documentation**:
23
+
24
+ * [Responsible Generative AI Toolkit](https://ai.google.dev/responsible)
25
+ * [Gemma on Kaggle](https://www.kaggle.com/models/google/gemma)
26
+ * [Gemma on Vertex Model Garden](https://console.cloud.google.com/vertex-ai/publishers/google/model-garden/335?version=gemma-7b-gg-hf)
27
+
28
+ **Terms of Use**: [Terms](https://www.kaggle.com/models/google/gemma/license/consent)
29
+
30
+ **Authors**: Google
31
+
32
+ ## Model Information
33
+
34
+ Summary description and brief definition of inputs and outputs.
35
+
36
+ ### Description
37
+
38
+ Gemma is a family of lightweight, state-of-the-art open models from Google,
39
+ built from the same research and technology used to create the Gemini models.
40
+ They are text-to-text, decoder-only large language models, available in English,
41
+ with open weights, pre-trained variants, and instruction-tuned variants. Gemma
42
+ models are well-suited for a variety of text generation tasks, including
43
+ question answering, summarization, and reasoning. Their relatively small size
44
+ makes it possible to deploy them in environments with limited resources such as
45
+ a laptop, desktop or your own cloud infrastructure, democratizing access to
46
+ state of the art AI models and helping foster innovation for everyone.
47
+
48
+ ### Usage
49
+
50
+ 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.
51
+
52
+ #### Fine-tuning examples
53
+
54
+ You can find fine-tuning notebooks under the [`examples/` directory](https://huggingface.co/google/gemma-7b/tree/main/examples). We provide:
55
+
56
+ * A script to perform Supervised Fine-Tuning (SFT) on UltraChat dataset using [QLoRA](https://huggingface.co/papers/2305.14314)
57
+ * A script to perform SFT using FSDP on TPU devices
58
+ * A notebook that you can run on a free-tier Google Colab instance to perform SFT on English quotes dataset. You can also find the copy of the notebook [here](https://github.com/huggingface/notebooks/blob/main/peft/gemma_7b_english_quotes.ipynb).
59
+
60
+ #### Running the model on a CPU
61
+
62
+
63
+ ```python
64
+ from transformers import AutoTokenizer, AutoModelForCausalLM
65
+
66
+ tokenizer = AutoTokenizer.from_pretrained("google/gemma-7b")
67
+ model = AutoModelForCausalLM.from_pretrained("google/gemma-7b")
68
+
69
+ input_text = "Write me a poem about Machine Learning."
70
+ input_ids = tokenizer(input_text, return_tensors="pt")
71
+
72
+ outputs = model.generate(**input_ids)
73
+ print(tokenizer.decode(outputs[0]))
74
+ ```
75
+
76
+
77
+ #### Running the model on a single / multi GPU
78
+
79
+
80
+ ```python
81
+ # pip install accelerate
82
+ from transformers import AutoTokenizer, AutoModelForCausalLM
83
+
84
+ tokenizer = AutoTokenizer.from_pretrained("google/gemma-7b")
85
+ model = AutoModelForCausalLM.from_pretrained("google/gemma-7b", device_map="auto")
86
+
87
+ input_text = "Write me a poem about Machine Learning."
88
+ input_ids = tokenizer(input_text, return_tensors="pt").to("cuda")
89
+
90
+ outputs = model.generate(**input_ids)
91
+ print(tokenizer.decode(outputs[0]))
92
+ ```
93
+
94
+
95
+ #### Running the model on a GPU using different precisions
96
+
97
+ * _Using `torch.float16`_
98
+
99
+ ```python
100
+ # pip install accelerate
101
+ from transformers import AutoTokenizer, AutoModelForCausalLM
102
+
103
+ tokenizer = AutoTokenizer.from_pretrained("google/gemma-7b")
104
+ model = AutoModelForCausalLM.from_pretrained("google/gemma-7b", device_map="auto", torch_dtype=torch.float16)
105
+
106
+ input_text = "Write me a poem about Machine Learning."
107
+ input_ids = tokenizer(input_text, return_tensors="pt").to("cuda")
108
+
109
+ outputs = model.generate(**input_ids)
110
+ print(tokenizer.decode(outputs[0]))
111
+ ```
112
+
113
+ * _Using `torch.bfloat16`_
114
+
115
+ ```python
116
+ # pip install accelerate
117
+ from transformers import AutoTokenizer, AutoModelForCausalLM
118
+
119
+ tokenizer = AutoTokenizer.from_pretrained("google/gemma-7b")
120
+ model = AutoModelForCausalLM.from_pretrained("google/gemma-7b", device_map="auto", torch_dtype=torch.bfloat16)
121
+
122
+ input_text = "Write me a poem about Machine Learning."
123
+ input_ids = tokenizer(input_text, return_tensors="pt").to("cuda")
124
+
125
+ outputs = model.generate(**input_ids)
126
+ print(tokenizer.decode(outputs[0]))
127
+ ```
128
+
129
+ #### Quantized Versions through `bitsandbytes`
130
+
131
+ * _Using 8-bit precision (int8)_
132
+
133
+ ```python
134
+ # pip install bitsandbytes accelerate
135
+ from transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig
136
+
137
+ quantization_config = BitsAndBytesConfig(load_in_8bit=True)
138
+
139
+ tokenizer = AutoTokenizer.from_pretrained("google/gemma-7b")
140
+ model = AutoModelForCausalLM.from_pretrained("google/gemma-7b", quantization_config=quantization_config)
141
+
142
+ input_text = "Write me a poem about Machine Learning."
143
+ input_ids = tokenizer(input_text, return_tensors="pt").to("cuda")
144
+
145
+ outputs = model.generate(**input_ids)
146
+ print(tokenizer.decode(outputs[0]))
147
+ ```
148
+
149
+ * _Using 4-bit precision_
150
+
151
+ ```python
152
+ # pip install bitsandbytes accelerate
153
+ from transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig
154
+
155
+ quantization_config = BitsAndBytesConfig(load_in_4bit=True)
156
+
157
+ tokenizer = AutoTokenizer.from_pretrained("google/gemma-7b")
158
+ model = AutoModelForCausalLM.from_pretrained("google/gemma-7b", quantization_config=quantization_config)
159
+
160
+ input_text = "Write me a poem about Machine Learning."
161
+ input_ids = tokenizer(input_text, return_tensors="pt").to("cuda")
162
+
163
+ outputs = model.generate(**input_ids)
164
+ print(tokenizer.decode(outputs[0]))
165
+ ```
166
+
167
+
168
+ #### Other optimizations
169
+
170
+ * _Flash Attention 2_
171
+
172
+ First make sure to install `flash-attn` in your environment `pip install flash-attn`
173
+
174
+ ```diff
175
+ model = AutoModelForCausalLM.from_pretrained(
176
+ model_id,
177
+ torch_dtype=torch.float16,
178
+ + attn_implementation="flash_attention_2"
179
+ ).to(0)
180
+ ```
181
+
182
+ ### Inputs and outputs
183
+
184
+ * **Input:** Text string, such as a question, a prompt, or a document to be
185
+ summarized.
186
+ * **Output:** Generated English-language text in response to the input, such
187
+ as an answer to a question, or a summary of a document.
188
+
189
+ ## Model Data
190
+
191
+ Data used for model training and how the data was processed.
192
+
193
+ ### Training Dataset
194
+
195
+ These models were trained on a dataset of text data that includes a wide variety
196
+ of sources, totaling 6 trillion tokens. Here are the key components:
197
+
198
+ * Web Documents: A diverse collection of web text ensures the model is exposed
199
+ to a broad range of linguistic styles, topics, and vocabulary. Primarily
200
+ English-language content.
201
+ * Code: Exposing the model to code helps it to learn the syntax and patterns of
202
+ programming languages, which improves its ability to generate code or
203
+ understand code-related questions.
204
+ * Mathematics: Training on mathematical text helps the model learn logical
205
+ reasoning, symbolic representation, and to address mathematical queries.
206
+
207
+ The combination of these diverse data sources is crucial for training a powerful
208
+ language model that can handle a wide variety of different tasks and text
209
+ formats.
210
+
211
+ ### Data Preprocessing
212
+
213
+ Here are the key data cleaning and filtering methods applied to the training
214
+ data:
215
+
216
+ * CSAM Filtering: Rigorous CSAM (Child Sexual Abuse Material) filtering was
217
+ applied at multiple stages in the data preparation process to ensure the
218
+ exclusion of harmful and illegal content
219
+ * Sensitive Data Filtering: As part of making Gemma pre-trained models safe and
220
+ reliable, automated techniques were used to filter out certain personal
221
+ information and other sensitive data from training sets.
222
+ * Additional methods: Filtering based on content quality and safely in line with
223
+ [our policies](https://storage.googleapis.com/gweb-uniblog-publish-prod/documents/2023_Google_AI_Principles_Progress_Update.pdf#page=11).
224
+
225
+ ## Implementation Information
226
+
227
+ Details about the model internals.
228
+
229
+ ### Hardware
230
+
231
+ Gemma was trained using the latest generation of
232
+ [Tensor Processing Unit (TPU)](https://cloud.google.com/tpu/docs/intro-to-tpu) hardware (TPUv5e).
233
+
234
+ Training large language models requires significant computational power. TPUs,
235
+ designed specifically for matrix operations common in machine learning, offer
236
+ several advantages in this domain:
237
+
238
+ * Performance: TPUs are specifically designed to handle the massive computations
239
+ involved in training LLMs. They can speed up training considerably compared to
240
+ CPUs.
241
+ * Memory: TPUs often come with large amounts of high-bandwidth memory, allowing
242
+ for the handling of large models and batch sizes during training. This can
243
+ lead to better model quality.
244
+ * Scalability: TPU Pods (large clusters of TPUs) provide a scalable solution for
245
+ handling the growing complexity of large foundation models. You can distribute
246
+ training across multiple TPU devices for faster and more efficient processing.
247
+ * Cost-effectiveness: In many scenarios, TPUs can provide a more cost-effective
248
+ solution for training large models compared to CPU-based infrastructure,
249
+ especially when considering the time and resources saved due to faster
250
+ training.
251
+ * These advantages are aligned with
252
+ [Google's commitments to operate sustainably](https://sustainability.google/operating-sustainably/).
253
+
254
+ ### Software
255
+
256
+ Training was done using [JAX](https://github.com/google/jax) and [ML Pathways](https://blog.google/technology/ai/introducing-pathways-next-generation-ai-architecture).
257
+
258
+ JAX allows researchers to take advantage of the latest generation of hardware,
259
+ including TPUs, for faster and more efficient training of large models.
260
+
261
+ ML Pathways is Google's latest effort to build artificially intelligent systems
262
+ capable of generalizing across multiple tasks. This is specially suitable for
263
+ [foundation models](https://ai.google/discover/foundation-models/), including large language models like
264
+ these ones.
265
+
266
+ Together, JAX and ML Pathways are used as described in the
267
+ [paper about the Gemini family of models](https://arxiv.org/abs/2312.11805); "the 'single
268
+ controller' programming model of Jax and Pathways allows a single Python
269
+ process to orchestrate the entire training run, dramatically simplifying the
270
+ development workflow."
271
+
272
+ ## Evaluation
273
+
274
+ Model evaluation metrics and results.
275
+
276
+ ### Benchmark Results
277
+
278
+ These models were evaluated against a large collection of different datasets and
279
+ metrics to cover different aspects of text generation:
280
+
281
+ | Benchmark | Metric | 2B Params | 7B Params |
282
+ | ------------------------------ | ------------- | ----------- | --------- |
283
+ | [MMLU](https://arxiv.org/abs/2009.03300) | 5-shot, top-1 | 42.3 | 64.3 |
284
+ | [HellaSwag](https://arxiv.org/abs/1905.07830) | 0-shot |71.4 | 81.2 |
285
+ | [PIQA](https://arxiv.org/abs/1911.11641) | 0-shot | 77.3 | 81.2 |
286
+ | [SocialIQA](https://arxiv.org/abs/1904.09728) | 0-shot | 59.7 | 51.8 |
287
+ | [BooIQ](https://arxiv.org/abs/1905.10044) | 0-shot | 69.4 | 83.2 |
288
+ | [WinoGrande](https://arxiv.org/abs/1907.10641) | partial score | 65.4 | 72.3 |
289
+ | [CommonsenseQA](https://arxiv.org/abs/1811.00937) | 7-shot | 65.3 | 71.3 |
290
+ | [OpenBookQA](https://arxiv.org/abs/1809.02789) | | 47.8 | 52.8 |
291
+ | [ARC-e](https://arxiv.org/abs/1911.01547) | | 73.2 | 81.5 |
292
+ | [ARC-c](https://arxiv.org/abs/1911.01547) | | 42.1 | 53.2 |
293
+ | [TriviaQA](https://arxiv.org/abs/1705.03551) | 5-shot | 53.2 | 63.4 |
294
+ | [Natural Questions](https://github.com/google-research-datasets/natural-questions) | 5-shot | - | 23 |
295
+ | [HumanEval](https://arxiv.org/abs/2107.03374) | pass@1 | 22.0 | 32.3 |
296
+ | [MBPP](https://arxiv.org/abs/2108.07732) | 3-shot | 29.2 | 44.4 |
297
+ | [GSM8K](https://arxiv.org/abs/2110.14168) | maj@1 | 17.7 | 46.4 |
298
+ | [MATH](https://arxiv.org/abs/2108.07732) | 4-shot | 11.8 | 24.3 |
299
+ | [AGIEval](https://arxiv.org/abs/2304.06364) | | 24.2 | 41.7 |
300
+ | [BIG-Bench](https://arxiv.org/abs/2206.04615) | | 35.2 | 55.1 |
301
+ | ------------------------------ | ------------- | ----------- | --------- |
302
+ | **Average** | | **54.0** | **56.4** |
303
+
304
+ ## Ethics and Safety
305
+
306
+ Ethics and safety evaluation approach and results.
307
+
308
+ ### Evaluation Approach
309
+
310
+ Our evaluation methods include structured evaluations and internal red-teaming
311
+ testing of relevant content policies. Red-teaming was conducted by a number of
312
+ different teams, each with different goals and human evaluation metrics. These
313
+ models were evaluated against a number of different categories relevant to
314
+ ethics and safety, including:
315
+
316
+ * Text-to-Text Content Safety: Human evaluation on prompts covering safety
317
+ policies including child sexual abuse and exploitation, harassment, violence
318
+ and gore, and hate speech.
319
+ * Text-to-Text Representational Harms: Benchmark against relevant academic
320
+ datasets such as [WinoBias](https://arxiv.org/abs/1804.06876) and [BBQ Dataset](https://arxiv.org/abs/2110.08193v2).
321
+ * Memorization: Automated evaluation of memorization of training data, including
322
+ the risk of personally identifiable information exposure.
323
+ * Large-scale harm: Tests for "dangerous capabilities," such as chemical,
324
+ biological, radiological, and nuclear (CBRN) risks.
325
+
326
+ ### Evaluation Results
327
+
328
+ The results of ethics and safety evaluations are within acceptable thresholds
329
+ for meeting [internal policies](https://storage.googleapis.com/gweb-uniblog-publish-prod/documents/2023_Google_AI_Principles_Progress_Update.pdf#page=11) for categories such as child
330
+ safety, content safety, representational harms, memorization, large-scale harms.
331
+ On top of robust internal evaluations, the results of well known safety
332
+ benchmarks like BBQ, BOLD, Winogender, Winobias, RealToxicity, and TruthfulQA
333
+ are shown here.
334
+
335
+ | Benchmark | Metric | 2B Params | 7B Params |
336
+ | ------------------------------ | ------------- | ----------- | --------- |
337
+ | [RealToxicity](https://arxiv.org/abs/2009.11462) | average | 6.86 | 7.90 |
338
+ | [BOLD](https://arxiv.org/abs/2101.11718) | | 45.57 | 49.08 |
339
+ | [CrowS-Pairs](https://aclanthology.org/2020.emnlp-main.154/) | top-1 | 45.82 | 51.33 |
340
+ | [BBQ Ambig](https://arxiv.org/abs/2110.08193v2) | 1-shot, top-1 | 62.58 | 92.54 |
341
+ | [BBQ Disambig](https://arxiv.org/abs/2110.08193v2) | top-1 | 54.62 | 71.99 |
342
+ | [Winogender](https://arxiv.org/abs/1804.09301) | top-1 | 51.25 | 54.17 |
343
+ | [TruthfulQA](https://arxiv.org/abs/2109.07958) | | 44.84 | 31.81 |
344
+ | [Winobias 1_2](https://arxiv.org/abs/1804.06876) | | 56.12 | 59.09 |
345
+ | [Winobias 2_2](https://arxiv.org/abs/1804.06876) | | 91.10 | 92.23 |
346
+ | [Toxigen](https://arxiv.org/abs/2203.09509) | | 29.77 | 39.59 |
347
+ | ------------------------------ | ------------- | ----------- | --------- |
348
+
349
+
350
+ ## Usage and Limitations
351
+
352
+ These models have certain limitations that users should be aware of.
353
+
354
+ ### Intended Usage
355
+
356
+ Open Large Language Models (LLMs) have a wide range of applications across
357
+ various industries and domains. The following list of potential uses is not
358
+ comprehensive. The purpose of this list is to provide contextual information
359
+ about the possible use-cases that the model creators considered as part of model
360
+ training and development.
361
+
362
+ * Content Creation and Communication
363
+ * Text Generation: These models can be used to generate creative text formats
364
+ such as poems, scripts, code, marketing copy, and email drafts.
365
+ * Chatbots and Conversational AI: Power conversational interfaces for customer
366
+ service, virtual assistants, or interactive applications.
367
+ * Text Summarization: Generate concise summaries of a text corpus, research
368
+ papers, or reports.
369
+ * Research and Education
370
+ * Natural Language Processing (NLP) Research: These models can serve as a
371
+ foundation for researchers to experiment with NLP techniques, develop
372
+ algorithms, and contribute to the advancement of the field.
373
+ * Language Learning Tools: Support interactive language learning experiences,
374
+ aiding in grammar correction or providing writing practice.
375
+ * Knowledge Exploration: Assist researchers in exploring large bodies of text
376
+ by generating summaries or answering questions about specific topics.
377
+
378
+ ### Limitations
379
+
380
+ * Training Data
381
+ * The quality and diversity of the training data significantly influence the
382
+ model's capabilities. Biases or gaps in the training data can lead to
383
+ limitations in the model's responses.
384
+ * The scope of the training dataset determines the subject areas the model can
385
+ handle effectively.
386
+ * Context and Task Complexity
387
+ * LLMs are better at tasks that can be framed with clear prompts and
388
+ instructions. Open-ended or highly complex tasks might be challenging.
389
+ * A model's performance can be influenced by the amount of context provided
390
+ (longer context generally leads to better outputs, up to a certain point).
391
+ * Language Ambiguity and Nuance
392
+ * Natural language is inherently complex. LLMs might struggle to grasp subtle
393
+ nuances, sarcasm, or figurative language.
394
+ * Factual Accuracy
395
+ * LLMs generate responses based on information they learned from their
396
+ training datasets, but they are not knowledge bases. They may generate
397
+ incorrect or outdated factual statements.
398
+ * Common Sense
399
+ * LLMs rely on statistical patterns in language. They might lack the ability
400
+ to apply common sense reasoning in certain situations.
401
+
402
+ ### Ethical Considerations and Risks
403
+
404
+ The development of large language models (LLMs) raises several ethical concerns.
405
+ In creating an open model, we have carefully considered the following:
406
+
407
+ * Bias and Fairness
408
+ * LLMs trained on large-scale, real-world text data can reflect socio-cultural
409
+ biases embedded in the training material. These models underwent careful
410
+ scrutiny, input data pre-processing described and posterior evaluations
411
+ reported in this card.
412
+ * Misinformation and Misuse
413
+ * LLMs can be misused to generate text that is false, misleading, or harmful.
414
+ * Guidelines are provided for responsible use with the model, see the
415
+ [Responsible Generative AI Toolkit](http://ai.google.dev/gemma/responsible).
416
+ * Transparency and Accountability:
417
+ * This model card summarizes details on the models' architecture,
418
+ capabilities, limitations, and evaluation processes.
419
+ * A responsibly developed open model offers the opportunity to share
420
+ innovation by making LLM technology accessible to developers and researchers
421
+ across the AI ecosystem.
422
+
423
+ Risks identified and mitigations:
424
+
425
+ * Perpetuation of biases: It's encouraged to perform continuous monitoring
426
+ (using evaluation metrics, human review) and the exploration of de-biasing
427
+ techniques during model training, fine-tuning, and other use cases.
428
+ * Generation of harmful content: Mechanisms and guidelines for content safety
429
+ are essential. Developers are encouraged to exercise caution and implement
430
+ appropriate content safety safeguards based on their specific product policies
431
+ and application use cases.
432
+ * Misuse for malicious purposes: Technical limitations and developer and
433
+ end-user education can help mitigate against malicious applications of LLMs.
434
+ Educational resources and reporting mechanisms for users to flag misuse are
435
+ provided. Prohibited uses of Gemma models are outlined in the
436
+ [Gemma Prohibited Use Policy](https://ai.google.dev/gemma/prohibited_use_policy).
437
+ * Privacy violations: Models were trained on data filtered for removal of PII
438
+ (Personally Identifiable Information). Developers are encouraged to adhere to
439
+ privacy regulations with privacy-preserving techniques.
440
+
441
+ ### Benefits
442
+
443
+ At the time of release, this family of models provides high-performance open
444
+ large language model implementations designed from the ground up for Responsible
445
+ AI development compared to similarly sized models.
446
+
447
+ Using the benchmark evaluation metrics described in this document, these models
448
+ have shown to provide superior performance to other, comparably-sized open model
449
+ alternatives.
450
+