camenduru commited on
Commit
91859f8
1 Parent(s): 371f3af

thanks to google gemma team ❤

Browse files
.gitattributes CHANGED
@@ -33,3 +33,5 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
33
  *.zip filter=lfs diff=lfs merge=lfs -text
34
  *.zst filter=lfs diff=lfs merge=lfs -text
35
  *tfevents* filter=lfs diff=lfs merge=lfs -text
 
 
 
33
  *.zip filter=lfs diff=lfs merge=lfs -text
34
  *.zst filter=lfs diff=lfs merge=lfs -text
35
  *tfevents* filter=lfs diff=lfs merge=lfs -text
36
+ tokenizer.json filter=lfs diff=lfs merge=lfs -text
37
+ gemma-2b.gguf filter=lfs diff=lfs merge=lfs -text
README.md ADDED
@@ -0,0 +1,447 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
+ ---
8
+
9
+ # Gemma Model Card
10
+
11
+ **Model Page**: [Gemma](https://ai.google.dev/gemma/docs)
12
+
13
+ This model card corresponds to the 2B base version of the Gemma model. You can also visit the model card of the [7B base model](https://huggingface.co/google/gemma-7b), [7B instruct model](https://huggingface.co/google/gemma-7b-it), and [2B instruct model](https://huggingface.co/google/gemma-2b-it).
14
+
15
+ **Resources and Technical Documentation**:
16
+
17
+ * [Responsible Generative AI Toolkit](https://ai.google.dev/responsible)
18
+ * [Gemma on Kaggle](https://www.kaggle.com/models/google/gemma)
19
+ * [Gemma on Vertex Model Garden](https://console.cloud.google.com/vertex-ai/publishers/google/model-garden/335)
20
+
21
+ **Terms of Use**: [Terms](https://www.kaggle.com/models/google/gemma/license/consent)
22
+
23
+ **Authors**: Google
24
+
25
+ ## Model Information
26
+
27
+ Summary description and brief definition of inputs and outputs.
28
+
29
+ ### Description
30
+
31
+ Gemma is a family of lightweight, state-of-the-art open models from Google,
32
+ built from the same research and technology used to create the Gemini models.
33
+ They are text-to-text, decoder-only large language models, available in English,
34
+ with open weights, pre-trained variants, and instruction-tuned variants. Gemma
35
+ models are well-suited for a variety of text generation tasks, including
36
+ question answering, summarization, and reasoning. Their relatively small size
37
+ makes it possible to deploy them in environments with limited resources such as
38
+ a laptop, desktop or your own cloud infrastructure, democratizing access to
39
+ state of the art AI models and helping foster innovation for everyone.
40
+
41
+ ### Usage
42
+
43
+ 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.
44
+
45
+
46
+ #### Fine-tuning the model
47
+
48
+ 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-2b`.
49
+ In that repository, we provide:
50
+
51
+ * A script to perform Supervised Fine-Tuning (SFT) on UltraChat dataset using QLoRA
52
+ * A script to perform SFT using FSDP on TPU devices
53
+ * A notebook that you can run on a free-tier Google Colab instance to perform SFT on English quotes dataset
54
+
55
+
56
+
57
+ #### Running the model on a CPU
58
+
59
+
60
+ ```python
61
+ from transformers import AutoTokenizer, AutoModelForCausalLM
62
+
63
+ tokenizer = AutoTokenizer.from_pretrained("google/gemma-2b")
64
+ model = AutoModelForCausalLM.from_pretrained("google/gemma-2b")
65
+
66
+ input_text = "Write me a poem about Machine Learning."
67
+ input_ids = tokenizer(**input_text, return_tensors="pt")
68
+
69
+ outputs = model.generate(input_ids)
70
+ print(tokenizer.decode(outputs[0]))
71
+ ```
72
+
73
+
74
+ #### Running the model on a single / multi GPU
75
+
76
+
77
+ ```python
78
+ # pip install accelerate
79
+ from transformers import AutoTokenizer, AutoModelForCausalLM
80
+
81
+ tokenizer = AutoTokenizer.from_pretrained("google/gemma-2b")
82
+ model = AutoModelForCausalLM.from_pretrained("google/gemma-2b", device_map="auto")
83
+
84
+ input_text = "Write me a poem about Machine Learning."
85
+ input_ids = tokenizer(input_text, return_tensors="pt").to("cuda")
86
+
87
+ outputs = model.generate(**input_ids)
88
+ print(tokenizer.decode(outputs[0]))
89
+ ```
90
+
91
+
92
+ #### Running the model on a GPU using different precisions
93
+
94
+ * _Using `torch.float16`_
95
+
96
+ ```python
97
+ # pip install accelerate
98
+ from transformers import AutoTokenizer, AutoModelForCausalLM
99
+
100
+ tokenizer = AutoTokenizer.from_pretrained("google/gemma-2b")
101
+ model = AutoModelForCausalLM.from_pretrained("google/gemma-2b", device_map="auto", torch_dtype=torch.float16)
102
+
103
+ input_text = "Write me a poem about Machine Learning."
104
+ input_ids = tokenizer(input_text, return_tensors="pt").to("cuda")
105
+
106
+ outputs = model.generate(**input_ids)
107
+ print(tokenizer.decode(outputs[0]))
108
+ ```
109
+
110
+ * _Using `torch.bfloat16`_
111
+
112
+ ```python
113
+ # pip install accelerate
114
+ from transformers import AutoTokenizer, AutoModelForCausalLM
115
+
116
+ tokenizer = AutoTokenizer.from_pretrained("google/gemma-2b")
117
+ model = AutoModelForCausalLM.from_pretrained("google/gemma-2b", device_map="auto", torch_dtype=torch.bfloat16)
118
+
119
+ input_text = "Write me a poem about Machine Learning."
120
+ input_ids = tokenizer(input_text, return_tensors="pt").to("cuda")
121
+
122
+ outputs = model.generate(**input_ids)
123
+ print(tokenizer.decode(outputs[0]))
124
+ ```
125
+
126
+ #### Quantized Versions through `bitsandbytes`
127
+
128
+ * _Using 8-bit precision (int8)_
129
+
130
+ ```python
131
+ # pip install bitsandbytes accelerate
132
+ from transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig
133
+
134
+ quantization_config = BitsAndBytesConfig(load_in_8bit=True)
135
+
136
+ tokenizer = AutoTokenizer.from_pretrained("google/gemma-2b")
137
+ model = AutoModelForCausalLM.from_pretrained("google/gemma-2b", quantization_config=quantization_config)
138
+
139
+ input_text = "Write me a poem about Machine Learning."
140
+ input_ids = tokenizer(input_text, return_tensors="pt").to("cuda")
141
+
142
+ outputs = model.generate(**input_ids)
143
+ print(tokenizer.decode(outputs[0]))
144
+ ```
145
+
146
+ * _Using 4-bit precision_
147
+
148
+ ```python
149
+ # pip install bitsandbytes accelerate
150
+ from transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig
151
+
152
+ quantization_config = BitsAndBytesConfig(load_in_4bit=True)
153
+
154
+ tokenizer = AutoTokenizer.from_pretrained("google/gemma-2b")
155
+ model = AutoModelForCausalLM.from_pretrained("google/gemma-2b", quantization_config=quantization_config)
156
+
157
+ input_text = "Write me a poem about Machine Learning."
158
+ input_ids = tokenizer(input_text, return_tensors="pt").to("cuda")
159
+
160
+ outputs = model.generate(**input_ids)
161
+ print(tokenizer.decode(outputs[0]))
162
+ ```
163
+
164
+
165
+ #### Other optimizations
166
+
167
+ * _Flash Attention 2_
168
+
169
+ First make sure to install `flash-attn` in your environment `pip install flash-attn`
170
+
171
+ ```diff
172
+ model = AutoModelForCausalLM.from_pretrained(
173
+ model_id,
174
+ torch_dtype=torch.float16,
175
+ + attn_implementation="flash_attention_2"
176
+ ).to(0)
177
+ ```
178
+
179
+ ### Inputs and outputs
180
+
181
+ * **Input:** Text string, such as a question, a prompt, or a document to be
182
+ summarized.
183
+ * **Output:** Generated English-language text in response to the input, such
184
+ as an answer to a question, or a summary of a document.
185
+
186
+ ## Model Data
187
+
188
+ Data used for model training and how the data was processed.
189
+
190
+ ### Training Dataset
191
+
192
+ These models were trained on a dataset of text data that includes a wide variety
193
+ of sources, totaling 6 trillion tokens. Here are the key components:
194
+
195
+ * Web Documents: A diverse collection of web text ensures the model is exposed
196
+ to a broad range of linguistic styles, topics, and vocabulary. Primarily
197
+ English-language content.
198
+ * Code: Exposing the model to code helps it to learn the syntax and patterns of
199
+ programming languages, which improves its ability to generate code or
200
+ understand code-related questions.
201
+ * Mathematics: Training on mathematical text helps the model learn logical
202
+ reasoning, symbolic representation, and to address mathematical queries.
203
+
204
+ The combination of these diverse data sources is crucial for training a powerful
205
+ language model that can handle a wide variety of different tasks and text
206
+ formats.
207
+
208
+ ### Data Preprocessing
209
+
210
+ Here are the key data cleaning and filtering methods applied to the training
211
+ data:
212
+
213
+ * CSAM Filtering: Rigorous CSAM (Child Sexual Abuse Material) filtering was
214
+ applied at multiple stages in the data preparation process to ensure the
215
+ exclusion of harmful and illegal content
216
+ * Sensitive Data Filtering: As part of making Gemma pre-trained models safe and
217
+ reliable, automated techniques were used to filter out certain personal
218
+ information and other sensitive data from training sets.
219
+ * Additional methods: Filtering based on content quality and safely in line with
220
+ [our policies](https://storage.googleapis.com/gweb-uniblog-publish-prod/documents/2023_Google_AI_Principles_Progress_Update.pdf#page=11).
221
+
222
+ ## Implementation Information
223
+
224
+ Details about the model internals.
225
+
226
+ ### Hardware
227
+
228
+ Gemma was trained using the latest generation of
229
+ [Tensor Processing Unit (TPU)](https://cloud.google.com/tpu/docs/intro-to-tpu) hardware (TPUv5e).
230
+
231
+ Training large language models requires significant computational power. TPUs,
232
+ designed specifically for matrix operations common in machine learning, offer
233
+ several advantages in this domain:
234
+
235
+ * Performance: TPUs are specifically designed to handle the massive computations
236
+ involved in training LLMs. They can speed up training considerably compared to
237
+ CPUs.
238
+ * Memory: TPUs often come with large amounts of high-bandwidth memory, allowing
239
+ for the handling of large models and batch sizes during training. This can
240
+ lead to better model quality.
241
+ * Scalability: TPU Pods (large clusters of TPUs) provide a scalable solution for
242
+ handling the growing complexity of large foundation models. You can distribute
243
+ training across multiple TPU devices for faster and more efficient processing.
244
+ * Cost-effectiveness: In many scenarios, TPUs can provide a more cost-effective
245
+ solution for training large models compared to CPU-based infrastructure,
246
+ especially when considering the time and resources saved due to faster
247
+ training.
248
+ * These advantages are aligned with
249
+ [Google's commitments to operate sustainably](https://sustainability.google/operating-sustainably/).
250
+
251
+ ### Software
252
+
253
+ Training was done using [JAX](https://github.com/google/jax) and [ML Pathways](https://blog.google/technology/ai/introducing-pathways-next-generation-ai-architecture/ml-pathways).
254
+
255
+ JAX allows researchers to take advantage of the latest generation of hardware,
256
+ including TPUs, for faster and more efficient training of large models.
257
+
258
+ ML Pathways is Google's latest effort to build artificially intelligent systems
259
+ capable of generalizing across multiple tasks. This is specially suitable for
260
+ [foundation models](https://ai.google/discover/foundation-models/), including large language models like
261
+ these ones.
262
+
263
+ Together, JAX and ML Pathways are used as described in the
264
+ [paper about the Gemini family of models](https://arxiv.org/abs/2312.11805); "the 'single
265
+ controller' programming model of Jax and Pathways allows a single Python
266
+ process to orchestrate the entire training run, dramatically simplifying the
267
+ development workflow."
268
+
269
+ ## Evaluation
270
+
271
+ Model evaluation metrics and results.
272
+
273
+ ### Benchmark Results
274
+
275
+ These models were evaluated against a large collection of different datasets and
276
+ metrics to cover different aspects of text generation:
277
+
278
+ | Benchmark | Metric | 2B Params | 7B Params |
279
+ | ------------------------------ | ------------- | ----------- | --------- |
280
+ | [MMLU](https://arxiv.org/abs/2009.03300) | 5-shot, top-1 | 42.3 | 64.3 |
281
+ | [HellaSwag](https://arxiv.org/abs/1905.07830) | 0-shot |71.4 | 81.2 |
282
+ | [PIQA](https://arxiv.org/abs/1911.11641) | 0-shot | 77.3 | 81.2 |
283
+ | [SocialIQA](https://arxiv.org/abs/1904.09728) | 0-shot | 59.7 | 51.8 |
284
+ | [BooIQ](https://arxiv.org/abs/1905.10044) | 0-shot | 69.4 | 83.2 |
285
+ | [WinoGrande](https://arxiv.org/abs/1907.10641) | partial score | 65.4 | 72.3 |
286
+ | [CommonsenseQA](https://arxiv.org/abs/1811.00937) | 7-shot | 65.3 | 71.3 |
287
+ | [OpenBookQA](https://arxiv.org/abs/1809.02789) | | 47.8 | 52.8 |
288
+ | [ARC-e](https://arxiv.org/abs/1911.01547) | | 73.2 | 81.5 |
289
+ | [ARC-c](https://arxiv.org/abs/1911.01547) | | 42.1 | 53.2 |
290
+ | [TriviaQA](https://arxiv.org/abs/1705.03551) | 5-shot | 53.2 | 63.4 |
291
+ | [Natural Questions](https://github.com/google-research-datasets/natural-questions) | 5-shot | - | 23 |
292
+ | [HumanEval](https://arxiv.org/abs/2107.03374) | pass@1 | 22.0 | 32.3 |
293
+ | [MBPP](https://arxiv.org/abs/2108.07732) | 3-shot | 29.2 | 44.4 |
294
+ | [GSM8K](https://arxiv.org/abs/2110.14168) | maj@1 | 17.7 | 46.4 |
295
+ | [MATH](https://arxiv.org/abs/2108.07732) | 4-shot | 11.8 | 24.3 |
296
+ | [AGIEval](https://arxiv.org/abs/2304.06364) | | 24.2 | 41.7 |
297
+ | [BIG-Bench](https://arxiv.org/abs/2206.04615) | | 35.2 | 55.1 |
298
+ | ------------------------------ | ------------- | ----------- | --------- |
299
+ | **Average** | | **54.0** | **56.4** |
300
+
301
+ ## Ethics and Safety
302
+
303
+ Ethics and safety evaluation approach and results.
304
+
305
+ ### Evaluation Approach
306
+
307
+ Our evaluation methods include structured evaluations and internal red-teaming
308
+ testing of relevant content policies. Red-teaming was conducted by a number of
309
+ different teams, each with different goals and human evaluation metrics. These
310
+ models were evaluated against a number of different categories relevant to
311
+ ethics and safety, including:
312
+
313
+ * Text-to-Text Content Safety: Human evaluation on prompts covering safety
314
+ policies including child sexual abuse and exploitation, harassment, violence
315
+ and gore, and hate speech.
316
+ * Text-to-Text Representational Harms: Benchmark against relevant academic
317
+ datasets such as [WinoBias](https://arxiv.org/abs/1804.06876) and [BBQ Dataset](https://arxiv.org/abs/2110.08193v2).
318
+ * Memorization: Automated evaluation of memorization of training data, including
319
+ the risk of personally identifiable information exposure.
320
+ * Large-scale harm: Tests for "dangerous capabilities," such as chemical,
321
+ biological, radiological, and nuclear (CBRN) risks.
322
+
323
+ ### Evaluation Results
324
+
325
+ The results of ethics and safety evaluations are within acceptable thresholds
326
+ 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
327
+ safety, content safety, representational harms, memorization, large-scale harms.
328
+ On top of robust internal evaluations, the results of well known safety
329
+ benchmarks like BBQ, BOLD, Winogender, Winobias, RealToxicity, and TruthfulQA
330
+ are shown here.
331
+
332
+ | Benchmark | Metric | 2B Params | 7B Params |
333
+ | ------------------------------ | ------------- | ----------- | --------- |
334
+ | [RealToxicity](https://arxiv.org/abs/2009.11462) | average | 6.86 | 7.90 |
335
+ | [BOLD](https://arxiv.org/abs/2101.11718) | | 45.57 | 49.08 |
336
+ | [CrowS-Pairs](https://aclanthology.org/2020.emnlp-main.154/) | top-1 | 45.82 | 51.33 |
337
+ | [BBQ Ambig](https://arxiv.org/abs/2110.08193v2) | 1-shot, top-1 | 62.58 | 92.54 |
338
+ | [BBQ Disambig](https://arxiv.org/abs/2110.08193v2) | top-1 | 54.62 | 71.99 |
339
+ | [Winogender](https://arxiv.org/abs/1804.09301) | top-1 | 51.25 | 54.17 |
340
+ | [TruthfulQA](https://arxiv.org/abs/2109.07958) | | 44.84 | 31.81 |
341
+ | [Winobias 1_2](https://arxiv.org/abs/1804.06876) | | 56.12 | 59.09 |
342
+ | [Winobias 2_2](https://arxiv.org/abs/1804.06876) | | 91.10 | 92.23 |
343
+ | [Toxigen](https://arxiv.org/abs/2203.09509) | | 29.77 | 39.59 |
344
+ | ------------------------------ | ------------- | ----------- | --------- |
345
+
346
+
347
+ ## Usage and Limitations
348
+
349
+ These models have certain limitations that users should be aware of.
350
+
351
+ ### Intended Usage
352
+
353
+ Open Large Language Models (LLMs) have a wide range of applications across
354
+ various industries and domains. The following list of potential uses is not
355
+ comprehensive. The purpose of this list is to provide contextual information
356
+ about the possible use-cases that the model creators considered as part of model
357
+ training and development.
358
+
359
+ * Content Creation and Communication
360
+ * Text Generation: These models can be used to generate creative text formats
361
+ such as poems, scripts, code, marketing copy, and email drafts.
362
+ * Chatbots and Conversational AI: Power conversational interfaces for customer
363
+ service, virtual assistants, or interactive applications.
364
+ * Text Summarization: Generate concise summaries of a text corpus, research
365
+ papers, or reports.
366
+ * Research and Education
367
+ * Natural Language Processing (NLP) Research: These models can serve as a
368
+ foundation for researchers to experiment with NLP techniques, develop
369
+ algorithms, and contribute to the advancement of the field.
370
+ * Language Learning Tools: Support interactive language learning experiences,
371
+ aiding in grammar correction or providing writing practice.
372
+ * Knowledge Exploration: Assist researchers in exploring large bodies of text
373
+ by generating summaries or answering questions about specific topics.
374
+
375
+ ### Limitations
376
+
377
+ * Training Data
378
+ * The quality and diversity of the training data significantly influence the
379
+ model's capabilities. Biases or gaps in the training data can lead to
380
+ limitations in the model's responses.
381
+ * The scope of the training dataset determines the subject areas the model can
382
+ handle effectively.
383
+ * Context and Task Complexity
384
+ * LLMs are better at tasks that can be framed with clear prompts and
385
+ instructions. Open-ended or highly complex tasks might be challenging.
386
+ * A model's performance can be influenced by the amount of context provided
387
+ (longer context generally leads to better outputs, up to a certain point).
388
+ * Language Ambiguity and Nuance
389
+ * Natural language is inherently complex. LLMs might struggle to grasp subtle
390
+ nuances, sarcasm, or figurative language.
391
+ * Factual Accuracy
392
+ * LLMs generate responses based on information they learned from their
393
+ training datasets, but they are not knowledge bases. They may generate
394
+ incorrect or outdated factual statements.
395
+ * Common Sense
396
+ * LLMs rely on statistical patterns in language. They might lack the ability
397
+ to apply common sense reasoning in certain situations.
398
+
399
+ ### Ethical Considerations and Risks
400
+
401
+ The development of large language models (LLMs) raises several ethical concerns.
402
+ In creating an open model, we have carefully considered the following:
403
+
404
+ * Bias and Fairness
405
+ * LLMs trained on large-scale, real-world text data can reflect socio-cultural
406
+ biases embedded in the training material. These models underwent careful
407
+ scrutiny, input data pre-processing described and posterior evaluations
408
+ reported in this card.
409
+ * Misinformation and Misuse
410
+ * LLMs can be misused to generate text that is false, misleading, or harmful.
411
+ * Guidelines are provided for responsible use with the model, see the
412
+ [Responsible Generative AI Toolkit](http://ai.google.dev/gemma/responsible).
413
+ * Transparency and Accountability:
414
+ * This model card summarizes details on the models' architecture,
415
+ capabilities, limitations, and evaluation processes.
416
+ * A responsibly developed open model offers the opportunity to share
417
+ innovation by making LLM technology accessible to developers and researchers
418
+ across the AI ecosystem.
419
+
420
+ Risks identified and mitigations:
421
+
422
+ * Perpetuation of biases: It's encouraged to perform continuous monitoring
423
+ (using evaluation metrics, human review) and the exploration of de-biasing
424
+ techniques during model training, fine-tuning, and other use cases.
425
+ * Generation of harmful content: Mechanisms and guidelines for content safety
426
+ are essential. Developers are encouraged to exercise caution and implement
427
+ appropriate content safety safeguards based on their specific product policies
428
+ and application use cases.
429
+ * Misuse for malicious purposes: Technical limitations and developer and
430
+ end-user education can help mitigate against malicious applications of LLMs.
431
+ Educational resources and reporting mechanisms for users to flag misuse are
432
+ provided. Prohibited uses of Gemma models are outlined in the
433
+ [Gemma Prohibited Use Policy](https://ai.google.dev/gemma/prohibited_use_policy).
434
+ * Privacy violations: Models were trained on data filtered for removal of PII
435
+ (Personally Identifiable Information). Developers are encouraged to adhere to
436
+ privacy regulations with privacy-preserving techniques.
437
+
438
+ ### Benefits
439
+
440
+ At the time of release, this family of models provides high-performance open
441
+ large language model implementations designed from the ground up for Responsible
442
+ AI development compared to similarly sized models.
443
+
444
+ Using the benchmark evaluation metrics described in this document, these models
445
+ have shown to provide superior performance to other, comparably-sized open model
446
+ alternatives.
447
+
config.json ADDED
@@ -0,0 +1,27 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "architectures": [
3
+ "GemmaForCausalLM"
4
+ ],
5
+ "attention_bias": false,
6
+ "attention_dropout": 0.0,
7
+ "bos_token_id": 2,
8
+ "eos_token_id": 1,
9
+ "head_dim": 256,
10
+ "hidden_act": "gelu",
11
+ "hidden_size": 2048,
12
+ "initializer_range": 0.02,
13
+ "intermediate_size": 16384,
14
+ "max_position_embeddings": 8192,
15
+ "model_type": "gemma",
16
+ "num_attention_heads": 8,
17
+ "num_hidden_layers": 18,
18
+ "num_key_value_heads": 1,
19
+ "pad_token_id": 0,
20
+ "rms_norm_eps": 1e-06,
21
+ "rope_scaling": null,
22
+ "rope_theta": 10000.0,
23
+ "torch_dtype": "bfloat16",
24
+ "transformers_version": "4.38.0.dev0",
25
+ "use_cache": true,
26
+ "vocab_size": 256000
27
+ }
generation_config.json ADDED
@@ -0,0 +1,7 @@
 
 
 
 
 
 
 
 
1
+ {
2
+ "_from_model_config": true,
3
+ "bos_token_id": 2,
4
+ "eos_token_id": 1,
5
+ "pad_token_id": 0,
6
+ "transformers_version": "4.38.0.dev0"
7
+ }
model-00001-of-00002.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:98caad07f1e1d17fede7734906d6881d4cb231e1d0cd2ebcf7fc8a6317eee4d6
3
+ size 4945242264
model-00002-of-00002.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:3342a2f7db857bc9873c699eef1416eb2b320603e5e042a6f7736689b26382cb
3
+ size 67121608
model.safetensors.index.json ADDED
@@ -0,0 +1,171 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "metadata": {
3
+ "total_size": 5012344832
4
+ },
5
+ "weight_map": {
6
+ "model.embed_tokens.weight": "model-00001-of-00002.safetensors",
7
+ "model.layers.0.input_layernorm.weight": "model-00001-of-00002.safetensors",
8
+ "model.layers.0.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
9
+ "model.layers.0.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
10
+ "model.layers.0.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
11
+ "model.layers.0.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
12
+ "model.layers.0.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
13
+ "model.layers.0.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
14
+ "model.layers.0.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
15
+ "model.layers.0.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
16
+ "model.layers.1.input_layernorm.weight": "model-00001-of-00002.safetensors",
17
+ "model.layers.1.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
18
+ "model.layers.1.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
19
+ "model.layers.1.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
20
+ "model.layers.1.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
21
+ "model.layers.1.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
22
+ "model.layers.1.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
23
+ "model.layers.1.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
24
+ "model.layers.1.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
25
+ "model.layers.10.input_layernorm.weight": "model-00001-of-00002.safetensors",
26
+ "model.layers.10.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
27
+ "model.layers.10.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
28
+ "model.layers.10.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
29
+ "model.layers.10.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
30
+ "model.layers.10.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
31
+ "model.layers.10.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
32
+ "model.layers.10.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
33
+ "model.layers.10.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
34
+ "model.layers.11.input_layernorm.weight": "model-00001-of-00002.safetensors",
35
+ "model.layers.11.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
36
+ "model.layers.11.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
37
+ "model.layers.11.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
38
+ "model.layers.11.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
39
+ "model.layers.11.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
40
+ "model.layers.11.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
41
+ "model.layers.11.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
42
+ "model.layers.11.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
43
+ "model.layers.12.input_layernorm.weight": "model-00001-of-00002.safetensors",
44
+ "model.layers.12.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
45
+ "model.layers.12.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
46
+ "model.layers.12.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
47
+ "model.layers.12.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
48
+ "model.layers.12.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
49
+ "model.layers.12.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
50
+ "model.layers.12.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
51
+ "model.layers.12.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
52
+ "model.layers.13.input_layernorm.weight": "model-00001-of-00002.safetensors",
53
+ "model.layers.13.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
54
+ "model.layers.13.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
55
+ "model.layers.13.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
56
+ "model.layers.13.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
57
+ "model.layers.13.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
58
+ "model.layers.13.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
59
+ "model.layers.13.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
60
+ "model.layers.13.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
61
+ "model.layers.14.input_layernorm.weight": "model-00001-of-00002.safetensors",
62
+ "model.layers.14.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
63
+ "model.layers.14.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
64
+ "model.layers.14.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
65
+ "model.layers.14.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
66
+ "model.layers.14.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
67
+ "model.layers.14.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
68
+ "model.layers.14.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
69
+ "model.layers.14.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
70
+ "model.layers.15.input_layernorm.weight": "model-00001-of-00002.safetensors",
71
+ "model.layers.15.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
72
+ "model.layers.15.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
73
+ "model.layers.15.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
74
+ "model.layers.15.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
75
+ "model.layers.15.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
76
+ "model.layers.15.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
77
+ "model.layers.15.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
78
+ "model.layers.15.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
79
+ "model.layers.16.input_layernorm.weight": "model-00001-of-00002.safetensors",
80
+ "model.layers.16.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
81
+ "model.layers.16.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
82
+ "model.layers.16.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
83
+ "model.layers.16.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
84
+ "model.layers.16.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
85
+ "model.layers.16.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
86
+ "model.layers.16.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
87
+ "model.layers.16.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
88
+ "model.layers.17.input_layernorm.weight": "model-00002-of-00002.safetensors",
89
+ "model.layers.17.mlp.down_proj.weight": "model-00002-of-00002.safetensors",
90
+ "model.layers.17.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
91
+ "model.layers.17.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
92
+ "model.layers.17.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
93
+ "model.layers.17.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
94
+ "model.layers.17.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
95
+ "model.layers.17.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
96
+ "model.layers.17.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
97
+ "model.layers.2.input_layernorm.weight": "model-00001-of-00002.safetensors",
98
+ "model.layers.2.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
99
+ "model.layers.2.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
100
+ "model.layers.2.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
101
+ "model.layers.2.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
102
+ "model.layers.2.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
103
+ "model.layers.2.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
104
+ "model.layers.2.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
105
+ "model.layers.2.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
106
+ "model.layers.3.input_layernorm.weight": "model-00001-of-00002.safetensors",
107
+ "model.layers.3.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
108
+ "model.layers.3.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
109
+ "model.layers.3.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
110
+ "model.layers.3.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
111
+ "model.layers.3.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
112
+ "model.layers.3.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
113
+ "model.layers.3.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
114
+ "model.layers.3.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
115
+ "model.layers.4.input_layernorm.weight": "model-00001-of-00002.safetensors",
116
+ "model.layers.4.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
117
+ "model.layers.4.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
118
+ "model.layers.4.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
119
+ "model.layers.4.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
120
+ "model.layers.4.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
121
+ "model.layers.4.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
122
+ "model.layers.4.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
123
+ "model.layers.4.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
124
+ "model.layers.5.input_layernorm.weight": "model-00001-of-00002.safetensors",
125
+ "model.layers.5.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
126
+ "model.layers.5.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
127
+ "model.layers.5.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
128
+ "model.layers.5.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
129
+ "model.layers.5.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
130
+ "model.layers.5.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
131
+ "model.layers.5.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
132
+ "model.layers.5.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
133
+ "model.layers.6.input_layernorm.weight": "model-00001-of-00002.safetensors",
134
+ "model.layers.6.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
135
+ "model.layers.6.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
136
+ "model.layers.6.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
137
+ "model.layers.6.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
138
+ "model.layers.6.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
139
+ "model.layers.6.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
140
+ "model.layers.6.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
141
+ "model.layers.6.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
142
+ "model.layers.7.input_layernorm.weight": "model-00001-of-00002.safetensors",
143
+ "model.layers.7.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
144
+ "model.layers.7.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
145
+ "model.layers.7.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
146
+ "model.layers.7.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
147
+ "model.layers.7.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
148
+ "model.layers.7.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
149
+ "model.layers.7.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
150
+ "model.layers.7.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
151
+ "model.layers.8.input_layernorm.weight": "model-00001-of-00002.safetensors",
152
+ "model.layers.8.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
153
+ "model.layers.8.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
154
+ "model.layers.8.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
155
+ "model.layers.8.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
156
+ "model.layers.8.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
157
+ "model.layers.8.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
158
+ "model.layers.8.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
159
+ "model.layers.8.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
160
+ "model.layers.9.input_layernorm.weight": "model-00001-of-00002.safetensors",
161
+ "model.layers.9.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
162
+ "model.layers.9.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
163
+ "model.layers.9.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
164
+ "model.layers.9.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
165
+ "model.layers.9.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
166
+ "model.layers.9.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
167
+ "model.layers.9.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
168
+ "model.layers.9.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
169
+ "model.norm.weight": "model-00002-of-00002.safetensors"
170
+ }
171
+ }
special_tokens_map.json ADDED
@@ -0,0 +1,30 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "bos_token": {
3
+ "content": "<bos>",
4
+ "lstrip": false,
5
+ "normalized": false,
6
+ "rstrip": false,
7
+ "single_word": false
8
+ },
9
+ "eos_token": {
10
+ "content": "<eos>",
11
+ "lstrip": false,
12
+ "normalized": false,
13
+ "rstrip": false,
14
+ "single_word": false
15
+ },
16
+ "pad_token": {
17
+ "content": "<pad>",
18
+ "lstrip": false,
19
+ "normalized": false,
20
+ "rstrip": false,
21
+ "single_word": false
22
+ },
23
+ "unk_token": {
24
+ "content": "<unk>",
25
+ "lstrip": false,
26
+ "normalized": false,
27
+ "rstrip": false,
28
+ "single_word": false
29
+ }
30
+ }
tokenizer.json ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d0d908b4f9326e0998815690e325b6abbd378978553e10627924dd825db7e243
3
+ size 17477553
tokenizer.model ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:6969e64047744a44bb3abfb5c50f8de0f7ed8b571d5444426ef931f651d1a0ef
3
+ size 4241111
tokenizer_config.json ADDED
@@ -0,0 +1,49 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "add_bos_token": true,
3
+ "add_eos_token": false,
4
+ "added_tokens_decoder": {
5
+ "0": {
6
+ "content": "<pad>",
7
+ "lstrip": false,
8
+ "normalized": false,
9
+ "rstrip": false,
10
+ "single_word": false,
11
+ "special": true
12
+ },
13
+ "1": {
14
+ "content": "<eos>",
15
+ "lstrip": false,
16
+ "normalized": false,
17
+ "rstrip": false,
18
+ "single_word": false,
19
+ "special": true
20
+ },
21
+ "2": {
22
+ "content": "<bos>",
23
+ "lstrip": false,
24
+ "normalized": false,
25
+ "rstrip": false,
26
+ "single_word": false,
27
+ "special": true
28
+ },
29
+ "3": {
30
+ "content": "<unk>",
31
+ "lstrip": false,
32
+ "normalized": false,
33
+ "rstrip": false,
34
+ "single_word": false,
35
+ "special": true
36
+ }
37
+ },
38
+ "bos_token": "<bos>",
39
+ "clean_up_tokenization_spaces": false,
40
+ "eos_token": "<eos>",
41
+ "legacy": null,
42
+ "model_max_length": 1000000000000000019884624838656,
43
+ "pad_token": "<pad>",
44
+ "sp_model_kwargs": {},
45
+ "spaces_between_special_tokens": false,
46
+ "tokenizer_class": "GemmaTokenizer",
47
+ "unk_token": "<unk>",
48
+ "use_default_system_prompt": false
49
+ }