TechxGenus commited on
Commit
8f760fc
1 Parent(s): 29b98a4

Upload README.md

Browse files
Files changed (1) hide show
  1. README.md +510 -0
README.md ADDED
@@ -0,0 +1,510 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ library_name: transformers
3
+ tags: []
4
+ widget:
5
+ - messages:
6
+ - role: user
7
+ content: How does the brain work?
8
+ inference:
9
+ parameters:
10
+ max_new_tokens: 200
11
+ extra_gated_heading: "Access Gemma on Hugging Face"
12
+ 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."
13
+ extra_gated_button_content: "Acknowledge license"
14
+ license: other
15
+ license_name: gemma-terms-of-use
16
+ license_link: https://ai.google.dev/gemma/terms
17
+ ---
18
+
19
+ GPTQ quantized version of gemma-7b-it model.
20
+
21
+ ---
22
+
23
+ # Gemma Model Card
24
+
25
+ **Model Page**: [Gemma](https://ai.google.dev/gemma/docs)
26
+
27
+ This model card corresponds to the 7B instruct version of the Gemma model. You can also visit the model card of the [2B base model](https://huggingface.co/google/gemma-2b), [7B base model](https://huggingface.co/google/gemma-7b), and [2B instruct model](https://huggingface.co/google/gemma-2b-it).
28
+
29
+ **Resources and Technical Documentation**:
30
+
31
+ * [Responsible Generative AI Toolkit](https://ai.google.dev/responsible)
32
+ * [Gemma on Kaggle](https://www.kaggle.com/models/google/gemma)
33
+ * [Gemma on Vertex Model Garden](https://console.cloud.google.com/vertex-ai/publishers/google/model-garden/335?version=gemma-7b-it-gg-hf)
34
+
35
+ **Terms of Use**: [Terms](https://www.kaggle.com/models/google/gemma/license/consent)
36
+
37
+ **Authors**: Google
38
+
39
+ ## Model Information
40
+
41
+ Summary description and brief definition of inputs and outputs.
42
+
43
+ ### Description
44
+
45
+ Gemma is a family of lightweight, state-of-the-art open models from Google,
46
+ built from the same research and technology used to create the Gemini models.
47
+ They are text-to-text, decoder-only large language models, available in English,
48
+ with open weights, pre-trained variants, and instruction-tuned variants. Gemma
49
+ models are well-suited for a variety of text generation tasks, including
50
+ question answering, summarization, and reasoning. Their relatively small size
51
+ makes it possible to deploy them in environments with limited resources such as
52
+ a laptop, desktop or your own cloud infrastructure, democratizing access to
53
+ state of the art AI models and helping foster innovation for everyone.
54
+
55
+ ### Usage
56
+
57
+ 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.
58
+
59
+ #### Fine-tuning the model
60
+
61
+ You can find fine-tuning scripts and notebook under the [`examples/` directory](https://huggingface.co/google/gemma-7b/tree/main/examples) of [`google/gemma-7b`](https://huggingface.co/google/gemma-7b) repository. To adapt it to this model, simply change the model-id to `google/gemma-7b-it`.
62
+ In that repository, we provide:
63
+
64
+ * A script to perform Supervised Fine-Tuning (SFT) on UltraChat dataset using QLoRA
65
+ * A script to perform SFT using FSDP on TPU devices
66
+ * A notebook that you can run on a free-tier Google Colab instance to perform SFT on English quotes dataset
67
+
68
+
69
+ #### Running the model on a CPU
70
+
71
+
72
+ ```python
73
+ from transformers import AutoTokenizer, AutoModelForCausalLM
74
+
75
+ tokenizer = AutoTokenizer.from_pretrained("google/gemma-7b-it")
76
+ model = AutoModelForCausalLM.from_pretrained("google/gemma-7b-it")
77
+
78
+ input_text = "Write me a poem about Machine Learning."
79
+ input_ids = tokenizer(input_text, return_tensors="pt")
80
+
81
+ outputs = model.generate(**input_ids)
82
+ print(tokenizer.decode(outputs[0]))
83
+ ```
84
+
85
+
86
+ #### Running the model on a single / multi GPU
87
+
88
+
89
+ ```python
90
+ # pip install accelerate
91
+ from transformers import AutoTokenizer, AutoModelForCausalLM
92
+
93
+ tokenizer = AutoTokenizer.from_pretrained("google/gemma-7b-it")
94
+ model = AutoModelForCausalLM.from_pretrained("google/gemma-7b-it", device_map="auto")
95
+
96
+ input_text = "Write me a poem about Machine Learning."
97
+ input_ids = tokenizer(input_text, return_tensors="pt").to("cuda")
98
+
99
+ outputs = model.generate(**input_ids)
100
+ print(tokenizer.decode(outputs[0]))
101
+ ```
102
+
103
+
104
+ #### Running the model on a GPU using different precisions
105
+
106
+ * _Using `torch.float16`_
107
+
108
+ ```python
109
+ # pip install accelerate
110
+ from transformers import AutoTokenizer, AutoModelForCausalLM
111
+
112
+ tokenizer = AutoTokenizer.from_pretrained("google/gemma-7b-it")
113
+ model = AutoModelForCausalLM.from_pretrained("google/gemma-7b-it", device_map="auto", torch_dtype=torch.float16)
114
+
115
+ input_text = "Write me a poem about Machine Learning."
116
+ input_ids = tokenizer(input_text, return_tensors="pt").to("cuda")
117
+
118
+ outputs = model.generate(**input_ids)
119
+ print(tokenizer.decode(outputs[0]))
120
+ ```
121
+
122
+ * _Using `torch.bfloat16`_
123
+
124
+ ```python
125
+ # pip install accelerate
126
+ from transformers import AutoTokenizer, AutoModelForCausalLM
127
+
128
+ tokenizer = AutoTokenizer.from_pretrained("google/gemma-7b-it")
129
+ model = AutoModelForCausalLM.from_pretrained("google/gemma-7b-it", device_map="auto", torch_dtype=torch.bfloat16)
130
+
131
+ input_text = "Write me a poem about Machine Learning."
132
+ input_ids = tokenizer(input_text, return_tensors="pt").to("cuda")
133
+
134
+ outputs = model.generate(**input_ids)
135
+ print(tokenizer.decode(outputs[0]))
136
+ ```
137
+
138
+ #### Quantized Versions through `bitsandbytes`
139
+
140
+ * _Using 8-bit precision (int8)_
141
+
142
+ ```python
143
+ # pip install bitsandbytes accelerate
144
+ from transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig
145
+
146
+ quantization_config = BitsAndBytesConfig(load_in_8bit=True)
147
+
148
+ tokenizer = AutoTokenizer.from_pretrained("google/gemma-7b-it")
149
+ model = AutoModelForCausalLM.from_pretrained("google/gemma-7b-it", quantization_config=quantization_config)
150
+
151
+ input_text = "Write me a poem about Machine Learning."
152
+ input_ids = tokenizer(input_text, return_tensors="pt").to("cuda")
153
+
154
+ outputs = model.generate(**input_ids)
155
+ print(tokenizer.decode(outputs[0]))
156
+ ```
157
+
158
+ * _Using 4-bit precision_
159
+
160
+ ```python
161
+ # pip install bitsandbytes accelerate
162
+ from transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig
163
+
164
+ quantization_config = BitsAndBytesConfig(load_in_4bit=True)
165
+
166
+ tokenizer = AutoTokenizer.from_pretrained("google/gemma-7b-it")
167
+ model = AutoModelForCausalLM.from_pretrained("google/gemma-7b-it", quantization_config=quantization_config)
168
+
169
+ input_text = "Write me a poem about Machine Learning."
170
+ input_ids = tokenizer(input_text, return_tensors="pt").to("cuda")
171
+
172
+ outputs = model.generate(**input_ids)
173
+ print(tokenizer.decode(outputs[0]))
174
+ ```
175
+
176
+
177
+ #### Other optimizations
178
+
179
+ * _Flash Attention 2_
180
+
181
+ First make sure to install `flash-attn` in your environment `pip install flash-attn`
182
+
183
+ ```diff
184
+ model = AutoModelForCausalLM.from_pretrained(
185
+ model_id,
186
+ torch_dtype=torch.float16,
187
+ + attn_implementation="flash_attention_2"
188
+ ).to(0)
189
+ ```
190
+
191
+ ### Chat Template
192
+
193
+ The instruction-tuned models use a chat template that must be adhered to for conversational use.
194
+ The easiest way to apply it is using the tokenizer's built-in chat template, as shown in the following snippet.
195
+
196
+ Let's load the model and apply the chat template to a conversation. In this example, we'll start with a single user interaction:
197
+
198
+ ```py
199
+ from transformers import AutoTokenizer, AutoModelForCausalLM
200
+ import transformers
201
+ import torch
202
+
203
+ model_id = "gg-hf/gemma-7b-it"
204
+ dtype = torch.bfloat16
205
+
206
+ tokenizer = AutoTokenizer.from_pretrained(model_id)
207
+ model = AutoModelForCausalLM.from_pretrained(
208
+ model_id,
209
+ device_map="cuda",
210
+ torch_dtype=dtype,
211
+ )
212
+
213
+ chat = [
214
+ { "role": "user", "content": "Write a hello world program" },
215
+ ]
216
+ prompt = tokenizer.apply_chat_template(chat, tokenize=False, add_generation_prompt=True)
217
+ ```
218
+
219
+ At this point, the prompt contains the following text:
220
+
221
+ ```
222
+ <bos><start_of_turn>user
223
+ Write a hello world program<end_of_turn>
224
+ <start_of_turn>model
225
+ ```
226
+
227
+ As you can see, each turn is preceded by a `<start_of_turn>` delimiter and then the role of the entity
228
+ (either `user`, for content supplied by the user, or `model` for LLM responses). Turns finish with
229
+ the `<end_of_turn>` token.
230
+
231
+ You can follow this format to build the prompt manually, if you need to do it without the tokenizer's
232
+ chat template.
233
+
234
+ After the prompt is ready, generation can be performed like this:
235
+
236
+ ```py
237
+ inputs = tokenizer.encode(prompt, add_special_tokens=False, return_tensors="pt")
238
+ outputs = model.generate(input_ids=inputs.to(model.device), max_new_tokens=150)
239
+ print(tokenizer.decode(outputs[0]))
240
+ ```
241
+
242
+ ### Inputs and outputs
243
+
244
+ * **Input:** Text string, such as a question, a prompt, or a document to be
245
+ summarized.
246
+ * **Output:** Generated English-language text in response to the input, such
247
+ as an answer to a question, or a summary of a document.
248
+
249
+ ## Model Data
250
+
251
+ Data used for model training and how the data was processed.
252
+
253
+ ### Training Dataset
254
+
255
+ These models were trained on a dataset of text data that includes a wide variety
256
+ of sources, totaling 6 trillion tokens. Here are the key components:
257
+
258
+ * Web Documents: A diverse collection of web text ensures the model is exposed
259
+ to a broad range of linguistic styles, topics, and vocabulary. Primarily
260
+ English-language content.
261
+ * Code: Exposing the model to code helps it to learn the syntax and patterns of
262
+ programming languages, which improves its ability to generate code or
263
+ understand code-related questions.
264
+ * Mathematics: Training on mathematical text helps the model learn logical
265
+ reasoning, symbolic representation, and to address mathematical queries.
266
+
267
+ The combination of these diverse data sources is crucial for training a powerful
268
+ language model that can handle a wide variety of different tasks and text
269
+ formats.
270
+
271
+ ### Data Preprocessing
272
+
273
+ Here are the key data cleaning and filtering methods applied to the training
274
+ data:
275
+
276
+ * CSAM Filtering: Rigorous CSAM (Child Sexual Abuse Material) filtering was
277
+ applied at multiple stages in the data preparation process to ensure the
278
+ exclusion of harmful and illegal content
279
+ * Sensitive Data Filtering: As part of making Gemma pre-trained models safe and
280
+ reliable, automated techniques were used to filter out certain personal
281
+ information and other sensitive data from training sets.
282
+ * Additional methods: Filtering based on content quality and safely in line with
283
+ [our policies](https://storage.googleapis.com/gweb-uniblog-publish-prod/documents/2023_Google_AI_Principles_Progress_Update.pdf#page=11).
284
+
285
+ ## Implementation Information
286
+
287
+ Details about the model internals.
288
+
289
+ ### Hardware
290
+
291
+ Gemma was trained using the latest generation of
292
+ [Tensor Processing Unit (TPU)](https://cloud.google.com/tpu/docs/intro-to-tpu) hardware (TPUv5e).
293
+
294
+ Training large language models requires significant computational power. TPUs,
295
+ designed specifically for matrix operations common in machine learning, offer
296
+ several advantages in this domain:
297
+
298
+ * Performance: TPUs are specifically designed to handle the massive computations
299
+ involved in training LLMs. They can speed up training considerably compared to
300
+ CPUs.
301
+ * Memory: TPUs often come with large amounts of high-bandwidth memory, allowing
302
+ for the handling of large models and batch sizes during training. This can
303
+ lead to better model quality.
304
+ * Scalability: TPU Pods (large clusters of TPUs) provide a scalable solution for
305
+ handling the growing complexity of large foundation models. You can distribute
306
+ training across multiple TPU devices for faster and more efficient processing.
307
+ * Cost-effectiveness: In many scenarios, TPUs can provide a more cost-effective
308
+ solution for training large models compared to CPU-based infrastructure,
309
+ especially when considering the time and resources saved due to faster
310
+ training.
311
+ * These advantages are aligned with
312
+ [Google's commitments to operate sustainably](https://sustainability.google/operating-sustainably/).
313
+
314
+ ### Software
315
+
316
+ Training was done using [JAX](https://github.com/google/jax) and [ML Pathways](https://blog.google/technology/ai/introducing-pathways-next-generation-ai-architecture).
317
+
318
+ JAX allows researchers to take advantage of the latest generation of hardware,
319
+ including TPUs, for faster and more efficient training of large models.
320
+
321
+ ML Pathways is Google's latest effort to build artificially intelligent systems
322
+ capable of generalizing across multiple tasks. This is specially suitable for
323
+ [foundation models](https://ai.google/discover/foundation-models/), including large language models like
324
+ these ones.
325
+
326
+ Together, JAX and ML Pathways are used as described in the
327
+ [paper about the Gemini family of models](https://arxiv.org/abs/2312.11805); "the 'single
328
+ controller' programming model of Jax and Pathways allows a single Python
329
+ process to orchestrate the entire training run, dramatically simplifying the
330
+ development workflow."
331
+
332
+ ## Evaluation
333
+
334
+ Model evaluation metrics and results.
335
+
336
+ ### Benchmark Results
337
+
338
+ These models were evaluated against a large collection of different datasets and
339
+ metrics to cover different aspects of text generation:
340
+
341
+ | Benchmark | Metric | 2B Params | 7B Params |
342
+ | ------------------------------ | ------------- | ----------- | --------- |
343
+ | [MMLU](https://arxiv.org/abs/2009.03300) | 5-shot, top-1 | 42.3 | 64.3 |
344
+ | [HellaSwag](https://arxiv.org/abs/1905.07830) | 0-shot |71.4 | 81.2 |
345
+ | [PIQA](https://arxiv.org/abs/1911.11641) | 0-shot | 77.3 | 81.2 |
346
+ | [SocialIQA](https://arxiv.org/abs/1904.09728) | 0-shot | 59.7 | 51.8 |
347
+ | [BooIQ](https://arxiv.org/abs/1905.10044) | 0-shot | 69.4 | 83.2 |
348
+ | [WinoGrande](https://arxiv.org/abs/1907.10641) | partial score | 65.4 | 72.3 |
349
+ | [CommonsenseQA](https://arxiv.org/abs/1811.00937) | 7-shot | 65.3 | 71.3 |
350
+ | [OpenBookQA](https://arxiv.org/abs/1809.02789) | | 47.8 | 52.8 |
351
+ | [ARC-e](https://arxiv.org/abs/1911.01547) | | 73.2 | 81.5 |
352
+ | [ARC-c](https://arxiv.org/abs/1911.01547) | | 42.1 | 53.2 |
353
+ | [TriviaQA](https://arxiv.org/abs/1705.03551) | 5-shot | 53.2 | 63.4 |
354
+ | [Natural Questions](https://github.com/google-research-datasets/natural-questions) | 5-shot | - | 23 |
355
+ | [HumanEval](https://arxiv.org/abs/2107.03374) | pass@1 | 22.0 | 32.3 |
356
+ | [MBPP](https://arxiv.org/abs/2108.07732) | 3-shot | 29.2 | 44.4 |
357
+ | [GSM8K](https://arxiv.org/abs/2110.14168) | maj@1 | 17.7 | 46.4 |
358
+ | [MATH](https://arxiv.org/abs/2108.07732) | 4-shot | 11.8 | 24.3 |
359
+ | [AGIEval](https://arxiv.org/abs/2304.06364) | | 24.2 | 41.7 |
360
+ | [BIG-Bench](https://arxiv.org/abs/2206.04615) | | 35.2 | 55.1 |
361
+ | ------------------------------ | ------------- | ----------- | --------- |
362
+ | **Average** | | **54.0** | **56.4** |
363
+
364
+ ## Ethics and Safety
365
+
366
+ Ethics and safety evaluation approach and results.
367
+
368
+ ### Evaluation Approach
369
+
370
+ Our evaluation methods include structured evaluations and internal red-teaming
371
+ testing of relevant content policies. Red-teaming was conducted by a number of
372
+ different teams, each with different goals and human evaluation metrics. These
373
+ models were evaluated against a number of different categories relevant to
374
+ ethics and safety, including:
375
+
376
+ * Text-to-Text Content Safety: Human evaluation on prompts covering safety
377
+ policies including child sexual abuse and exploitation, harassment, violence
378
+ and gore, and hate speech.
379
+ * Text-to-Text Representational Harms: Benchmark against relevant academic
380
+ datasets such as [WinoBias](https://arxiv.org/abs/1804.06876) and [BBQ Dataset](https://arxiv.org/abs/2110.08193v2).
381
+ * Memorization: Automated evaluation of memorization of training data, including
382
+ the risk of personally identifiable information exposure.
383
+ * Large-scale harm: Tests for "dangerous capabilities," such as chemical,
384
+ biological, radiological, and nuclear (CBRN) risks.
385
+
386
+ ### Evaluation Results
387
+
388
+ The results of ethics and safety evaluations are within acceptable thresholds
389
+ 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
390
+ safety, content safety, representational harms, memorization, large-scale harms.
391
+ On top of robust internal evaluations, the results of well known safety
392
+ benchmarks like BBQ, BOLD, Winogender, Winobias, RealToxicity, and TruthfulQA
393
+ are shown here.
394
+
395
+ | Benchmark | Metric | 2B Params | 7B Params |
396
+ | ------------------------------ | ------------- | ----------- | --------- |
397
+ | [RealToxicity](https://arxiv.org/abs/2009.11462) | average | 6.86 | 7.90 |
398
+ | [BOLD](https://arxiv.org/abs/2101.11718) | | 45.57 | 49.08 |
399
+ | [CrowS-Pairs](https://aclanthology.org/2020.emnlp-main.154/) | top-1 | 45.82 | 51.33 |
400
+ | [BBQ Ambig](https://arxiv.org/abs/2110.08193v2) | 1-shot, top-1 | 62.58 | 92.54 |
401
+ | [BBQ Disambig](https://arxiv.org/abs/2110.08193v2) | top-1 | 54.62 | 71.99 |
402
+ | [Winogender](https://arxiv.org/abs/1804.09301) | top-1 | 51.25 | 54.17 |
403
+ | [TruthfulQA](https://arxiv.org/abs/2109.07958) | | 44.84 | 31.81 |
404
+ | [Winobias 1_2](https://arxiv.org/abs/1804.06876) | | 56.12 | 59.09 |
405
+ | [Winobias 2_2](https://arxiv.org/abs/1804.06876) | | 91.10 | 92.23 |
406
+ | [Toxigen](https://arxiv.org/abs/2203.09509) | | 29.77 | 39.59 |
407
+ | ------------------------------ | ------------- | ----------- | --------- |
408
+
409
+
410
+ ## Usage and Limitations
411
+
412
+ These models have certain limitations that users should be aware of.
413
+
414
+ ### Intended Usage
415
+
416
+ Open Large Language Models (LLMs) have a wide range of applications across
417
+ various industries and domains. The following list of potential uses is not
418
+ comprehensive. The purpose of this list is to provide contextual information
419
+ about the possible use-cases that the model creators considered as part of model
420
+ training and development.
421
+
422
+ * Content Creation and Communication
423
+ * Text Generation: These models can be used to generate creative text formats
424
+ such as poems, scripts, code, marketing copy, and email drafts.
425
+ * Chatbots and Conversational AI: Power conversational interfaces for customer
426
+ service, virtual assistants, or interactive applications.
427
+ * Text Summarization: Generate concise summaries of a text corpus, research
428
+ papers, or reports.
429
+ * Research and Education
430
+ * Natural Language Processing (NLP) Research: These models can serve as a
431
+ foundation for researchers to experiment with NLP techniques, develop
432
+ algorithms, and contribute to the advancement of the field.
433
+ * Language Learning Tools: Support interactive language learning experiences,
434
+ aiding in grammar correction or providing writing practice.
435
+ * Knowledge Exploration: Assist researchers in exploring large bodies of text
436
+ by generating summaries or answering questions about specific topics.
437
+
438
+ ### Limitations
439
+
440
+ * Training Data
441
+ * The quality and diversity of the training data significantly influence the
442
+ model's capabilities. Biases or gaps in the training data can lead to
443
+ limitations in the model's responses.
444
+ * The scope of the training dataset determines the subject areas the model can
445
+ handle effectively.
446
+ * Context and Task Complexity
447
+ * LLMs are better at tasks that can be framed with clear prompts and
448
+ instructions. Open-ended or highly complex tasks might be challenging.
449
+ * A model's performance can be influenced by the amount of context provided
450
+ (longer context generally leads to better outputs, up to a certain point).
451
+ * Language Ambiguity and Nuance
452
+ * Natural language is inherently complex. LLMs might struggle to grasp subtle
453
+ nuances, sarcasm, or figurative language.
454
+ * Factual Accuracy
455
+ * LLMs generate responses based on information they learned from their
456
+ training datasets, but they are not knowledge bases. They may generate
457
+ incorrect or outdated factual statements.
458
+ * Common Sense
459
+ * LLMs rely on statistical patterns in language. They might lack the ability
460
+ to apply common sense reasoning in certain situations.
461
+
462
+ ### Ethical Considerations and Risks
463
+
464
+ The development of large language models (LLMs) raises several ethical concerns.
465
+ In creating an open model, we have carefully considered the following:
466
+
467
+ * Bias and Fairness
468
+ * LLMs trained on large-scale, real-world text data can reflect socio-cultural
469
+ biases embedded in the training material. These models underwent careful
470
+ scrutiny, input data pre-processing described and posterior evaluations
471
+ reported in this card.
472
+ * Misinformation and Misuse
473
+ * LLMs can be misused to generate text that is false, misleading, or harmful.
474
+ * Guidelines are provided for responsible use with the model, see the
475
+ [Responsible Generative AI Toolkit](http://ai.google.dev/gemma/responsible).
476
+ * Transparency and Accountability:
477
+ * This model card summarizes details on the models' architecture,
478
+ capabilities, limitations, and evaluation processes.
479
+ * A responsibly developed open model offers the opportunity to share
480
+ innovation by making LLM technology accessible to developers and researchers
481
+ across the AI ecosystem.
482
+
483
+ Risks identified and mitigations:
484
+
485
+ * Perpetuation of biases: It's encouraged to perform continuous monitoring
486
+ (using evaluation metrics, human review) and the exploration of de-biasing
487
+ techniques during model training, fine-tuning, and other use cases.
488
+ * Generation of harmful content: Mechanisms and guidelines for content safety
489
+ are essential. Developers are encouraged to exercise caution and implement
490
+ appropriate content safety safeguards based on their specific product policies
491
+ and application use cases.
492
+ * Misuse for malicious purposes: Technical limitations and developer and
493
+ end-user education can help mitigate against malicious applications of LLMs.
494
+ Educational resources and reporting mechanisms for users to flag misuse are
495
+ provided. Prohibited uses of Gemma models are outlined in the
496
+ [Gemma Prohibited Use Policy](https://ai.google.dev/gemma/prohibited_use_policy).
497
+ * Privacy violations: Models were trained on data filtered for removal of PII
498
+ (Personally Identifiable Information). Developers are encouraged to adhere to
499
+ privacy regulations with privacy-preserving techniques.
500
+
501
+ ### Benefits
502
+
503
+ At the time of release, this family of models provides high-performance open
504
+ large language model implementations designed from the ground up for Responsible
505
+ AI development compared to similarly sized models.
506
+
507
+ Using the benchmark evaluation metrics described in this document, these models
508
+ have shown to provide superior performance to other, comparably-sized open model
509
+ alternatives.
510
+