nold commited on
Commit
283de33
1 Parent(s): 31b2cf9

Upload folder using huggingface_hub (#1)

Browse files

- 7c517cb83ab92ef5cc0e183b708f3374a03327a6d1afe67599b3962610da38a5 (4c983abf8f6f25cc8d0167cd2a77187e9b03afcd)
- 6230660387be9e71eb0fb29a27c068ecab7514f37aed90bcfbd7ed991efe1418 (1c4a5ee73af1d20f190e88a441c88c3a2667ff8b)
- 7e78b383b207ce732ca3169c0f058dfbee9f0a70fd5a21fb6b39e5f7ec84d045 (29ae10584ff5c9d847351532d2c07007bd84e9ce)
- 88850c94a1056e0b1c23bbbe6bcd17ed2782625beacfdd8333f7f7783685a696 (4d00e5119d4d39e44f9a1b3b2c40c040440a1a6f)
- 11e6435a259c02e95a2a1131992e6e5568ef8d0588b618cd1a9638f12a947f08 (aa9ac4af5e8611b73a0fc616f1969d947a13e188)
- 9a1943d1681009dfb772064e3b1d38344f5c1a25ef295ce93fd5117538116e5a (aaa59a0aec183e2410bc5b41f25cb031ff179ac4)
- 6dc5fedb5f382e78c55fe1ebd99ad26d8201c4c57a17320f30499a9e8ee7dec9 (7905528a38513fbc486cfb5df1bf602276995ecb)
- 499883084fbfa8b59583c28c767d765b0fbdf41567dea098e0654a2d2bcc9808 (830d52b62116b25c8d00710c35992aec62ea5d51)

.gitattributes CHANGED
@@ -33,3 +33,12 @@ 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
+ gemma-2b_IQ3_S.gguf filter=lfs diff=lfs merge=lfs -text
37
+ gemma-2b_IQ4_NL.gguf filter=lfs diff=lfs merge=lfs -text
38
+ gemma-2b_Q2_K.gguf filter=lfs diff=lfs merge=lfs -text
39
+ gemma-2b_Q3_K_M.gguf filter=lfs diff=lfs merge=lfs -text
40
+ gemma-2b_Q4_K_M.gguf filter=lfs diff=lfs merge=lfs -text
41
+ gemma-2b_Q5_K_M.gguf filter=lfs diff=lfs merge=lfs -text
42
+ gemma-2b_Q8_0.gguf filter=lfs diff=lfs merge=lfs -text
43
+ gemma-2b_Q6_K.gguf filter=lfs diff=lfs merge=lfs -text
44
+ imatrix.dat filter=lfs diff=lfs merge=lfs -text
README.md ADDED
@@ -0,0 +1,460 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ library_name: transformers
3
+ extra_gated_heading: Access Gemma on Hugging Face
4
+ extra_gated_prompt: >-
5
+ To access Gemma on Hugging Face, you’re required to review and agree to
6
+ Google’s usage license. To do this, please ensure you’re logged-in to Hugging
7
+ Face and click below. Requests are processed immediately.
8
+ extra_gated_button_content: Acknowledge license
9
+ license: gemma
10
+ ---
11
+
12
+ # Gemma Model Card
13
+
14
+ **Model Page**: [Gemma](https://ai.google.dev/gemma/docs)
15
+
16
+ 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).
17
+
18
+ **Resources and Technical Documentation**:
19
+
20
+ * [Gemma Technical Report](https://storage.googleapis.com/deepmind-media/gemma/gemma-report.pdf)
21
+ * [Responsible Generative AI Toolkit](https://ai.google.dev/responsible)
22
+ * [Gemma on Kaggle](https://www.kaggle.com/models/google/gemma)
23
+ * [Gemma on Vertex Model Garden](https://console.cloud.google.com/vertex-ai/publishers/google/model-garden/335?version=gemma-2b-gg-hf)
24
+
25
+ **Terms of Use**: [Terms](https://www.kaggle.com/models/google/gemma/license/consent)
26
+
27
+ **Authors**: Google
28
+
29
+ ## Model Information
30
+
31
+ Summary description and brief definition of inputs and outputs.
32
+
33
+ ### Description
34
+
35
+ Gemma is a family of lightweight, state-of-the-art open models from Google,
36
+ built from the same research and technology used to create the Gemini models.
37
+ They are text-to-text, decoder-only large language models, available in English,
38
+ with open weights, pre-trained variants, and instruction-tuned variants. Gemma
39
+ models are well-suited for a variety of text generation tasks, including
40
+ question answering, summarization, and reasoning. Their relatively small size
41
+ makes it possible to deploy them in environments with limited resources such as
42
+ a laptop, desktop or your own cloud infrastructure, democratizing access to
43
+ state of the art AI models and helping foster innovation for everyone.
44
+
45
+ ### Context Length
46
+ Models are trained on a context length of 8192 tokens.
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
+
53
+ #### Fine-tuning the model
54
+
55
+ 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`.
56
+ In that repository, we provide:
57
+
58
+ * A script to perform Supervised Fine-Tuning (SFT) on UltraChat dataset using QLoRA
59
+ * A script to perform SFT using FSDP on TPU devices
60
+ * A notebook that you can run on a free-tier Google Colab instance to perform SFT on English quotes dataset
61
+
62
+
63
+
64
+ #### Running the model on a CPU
65
+
66
+
67
+ ```python
68
+ from transformers import AutoTokenizer, AutoModelForCausalLM
69
+
70
+ tokenizer = AutoTokenizer.from_pretrained("google/gemma-2b")
71
+ model = AutoModelForCausalLM.from_pretrained("google/gemma-2b")
72
+
73
+ input_text = "Write me a poem about Machine Learning."
74
+ input_ids = tokenizer(input_text, return_tensors="pt")
75
+
76
+ outputs = model.generate(**input_ids)
77
+ print(tokenizer.decode(outputs[0]))
78
+ ```
79
+
80
+
81
+ #### Running the model on a single / multi GPU
82
+
83
+
84
+ ```python
85
+ # pip install accelerate
86
+ from transformers import AutoTokenizer, AutoModelForCausalLM
87
+
88
+ tokenizer = AutoTokenizer.from_pretrained("google/gemma-2b")
89
+ model = AutoModelForCausalLM.from_pretrained("google/gemma-2b", device_map="auto")
90
+
91
+ input_text = "Write me a poem about Machine Learning."
92
+ input_ids = tokenizer(input_text, return_tensors="pt").to("cuda")
93
+
94
+ outputs = model.generate(**input_ids)
95
+ print(tokenizer.decode(outputs[0]))
96
+ ```
97
+
98
+
99
+ #### Running the model on a GPU using different precisions
100
+
101
+ * _Using `torch.float16`_
102
+
103
+ ```python
104
+ # pip install accelerate
105
+ from transformers import AutoTokenizer, AutoModelForCausalLM
106
+
107
+ tokenizer = AutoTokenizer.from_pretrained("google/gemma-2b")
108
+ model = AutoModelForCausalLM.from_pretrained("google/gemma-2b", device_map="auto", revision="float16")
109
+
110
+ input_text = "Write me a poem about Machine Learning."
111
+ input_ids = tokenizer(input_text, return_tensors="pt").to("cuda")
112
+
113
+ outputs = model.generate(**input_ids)
114
+ print(tokenizer.decode(outputs[0]))
115
+ ```
116
+
117
+ * _Using `torch.bfloat16`_
118
+
119
+ ```python
120
+ # pip install accelerate
121
+ from transformers import AutoTokenizer, AutoModelForCausalLM
122
+
123
+ tokenizer = AutoTokenizer.from_pretrained("google/gemma-2b")
124
+ model = AutoModelForCausalLM.from_pretrained("google/gemma-2b", device_map="auto", torch_dtype=torch.bfloat16)
125
+
126
+ input_text = "Write me a poem about Machine Learning."
127
+ input_ids = tokenizer(input_text, return_tensors="pt").to("cuda")
128
+
129
+ outputs = model.generate(**input_ids)
130
+ print(tokenizer.decode(outputs[0]))
131
+ ```
132
+
133
+ #### Quantized Versions through `bitsandbytes`
134
+
135
+ * _Using 8-bit precision (int8)_
136
+
137
+ ```python
138
+ # pip install bitsandbytes accelerate
139
+ from transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig
140
+
141
+ quantization_config = BitsAndBytesConfig(load_in_8bit=True)
142
+
143
+ tokenizer = AutoTokenizer.from_pretrained("google/gemma-2b")
144
+ model = AutoModelForCausalLM.from_pretrained("google/gemma-2b", quantization_config=quantization_config)
145
+
146
+ input_text = "Write me a poem about Machine Learning."
147
+ input_ids = tokenizer(input_text, return_tensors="pt").to("cuda")
148
+
149
+ outputs = model.generate(**input_ids)
150
+ print(tokenizer.decode(outputs[0]))
151
+ ```
152
+
153
+ * _Using 4-bit precision_
154
+
155
+ ```python
156
+ # pip install bitsandbytes accelerate
157
+ from transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig
158
+
159
+ quantization_config = BitsAndBytesConfig(load_in_4bit=True)
160
+
161
+ tokenizer = AutoTokenizer.from_pretrained("google/gemma-2b")
162
+ model = AutoModelForCausalLM.from_pretrained("google/gemma-2b", quantization_config=quantization_config)
163
+
164
+ input_text = "Write me a poem about Machine Learning."
165
+ input_ids = tokenizer(input_text, return_tensors="pt").to("cuda")
166
+
167
+ outputs = model.generate(**input_ids)
168
+ print(tokenizer.decode(outputs[0]))
169
+ ```
170
+
171
+
172
+ #### Other optimizations
173
+
174
+ * _Flash Attention 2_
175
+
176
+ First make sure to install `flash-attn` in your environment `pip install flash-attn`
177
+
178
+ ```diff
179
+ model = AutoModelForCausalLM.from_pretrained(
180
+ model_id,
181
+ torch_dtype=torch.float16,
182
+ + attn_implementation="flash_attention_2"
183
+ ).to(0)
184
+ ```
185
+
186
+ ### Inputs and outputs
187
+
188
+ * **Input:** Text string, such as a question, a prompt, or a document to be
189
+ summarized.
190
+ * **Output:** Generated English-language text in response to the input, such
191
+ as an answer to a question, or a summary of a document.
192
+
193
+ ## Model Data
194
+
195
+ Data used for model training and how the data was processed.
196
+
197
+ ### Training Dataset
198
+
199
+ These models were trained on a dataset of text data that includes a wide variety
200
+ of sources, totaling 6 trillion tokens. Here are the key components:
201
+
202
+ * Web Documents: A diverse collection of web text ensures the model is exposed
203
+ to a broad range of linguistic styles, topics, and vocabulary. Primarily
204
+ English-language content.
205
+ * Code: Exposing the model to code helps it to learn the syntax and patterns of
206
+ programming languages, which improves its ability to generate code or
207
+ understand code-related questions.
208
+ * Mathematics: Training on mathematical text helps the model learn logical
209
+ reasoning, symbolic representation, and to address mathematical queries.
210
+
211
+ The combination of these diverse data sources is crucial for training a powerful
212
+ language model that can handle a wide variety of different tasks and text
213
+ formats.
214
+
215
+ ### Data Preprocessing
216
+
217
+ Here are the key data cleaning and filtering methods applied to the training
218
+ data:
219
+
220
+ * CSAM Filtering: Rigorous CSAM (Child Sexual Abuse Material) filtering was
221
+ applied at multiple stages in the data preparation process to ensure the
222
+ exclusion of harmful and illegal content
223
+ * Sensitive Data Filtering: As part of making Gemma pre-trained models safe and
224
+ reliable, automated techniques were used to filter out certain personal
225
+ information and other sensitive data from training sets.
226
+ * Additional methods: Filtering based on content quality and safely in line with
227
+ [our policies](https://storage.googleapis.com/gweb-uniblog-publish-prod/documents/2023_Google_AI_Principles_Progress_Update.pdf#page=11).
228
+
229
+ ## Implementation Information
230
+
231
+ Details about the model internals.
232
+
233
+ ### Hardware
234
+
235
+ Gemma was trained using the latest generation of
236
+ [Tensor Processing Unit (TPU)](https://cloud.google.com/tpu/docs/intro-to-tpu) hardware (TPUv5e).
237
+
238
+ Training large language models requires significant computational power. TPUs,
239
+ designed specifically for matrix operations common in machine learning, offer
240
+ several advantages in this domain:
241
+
242
+ * Performance: TPUs are specifically designed to handle the massive computations
243
+ involved in training LLMs. They can speed up training considerably compared to
244
+ CPUs.
245
+ * Memory: TPUs often come with large amounts of high-bandwidth memory, allowing
246
+ for the handling of large models and batch sizes during training. This can
247
+ lead to better model quality.
248
+ * Scalability: TPU Pods (large clusters of TPUs) provide a scalable solution for
249
+ handling the growing complexity of large foundation models. You can distribute
250
+ training across multiple TPU devices for faster and more efficient processing.
251
+ * Cost-effectiveness: In many scenarios, TPUs can provide a more cost-effective
252
+ solution for training large models compared to CPU-based infrastructure,
253
+ especially when considering the time and resources saved due to faster
254
+ training.
255
+ * These advantages are aligned with
256
+ [Google's commitments to operate sustainably](https://sustainability.google/operating-sustainably/).
257
+
258
+ ### Software
259
+
260
+ 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).
261
+
262
+ JAX allows researchers to take advantage of the latest generation of hardware,
263
+ including TPUs, for faster and more efficient training of large models.
264
+
265
+ ML Pathways is Google's latest effort to build artificially intelligent systems
266
+ capable of generalizing across multiple tasks. This is specially suitable for
267
+ [foundation models](https://ai.google/discover/foundation-models/), including large language models like
268
+ these ones.
269
+
270
+ Together, JAX and ML Pathways are used as described in the
271
+ [paper about the Gemini family of models](https://arxiv.org/abs/2312.11805); "the 'single
272
+ controller' programming model of Jax and Pathways allows a single Python
273
+ process to orchestrate the entire training run, dramatically simplifying the
274
+ development workflow."
275
+
276
+ ## Evaluation
277
+
278
+ Model evaluation metrics and results.
279
+
280
+ ### Benchmark Results
281
+
282
+ These models were evaluated against a large collection of different datasets and
283
+ metrics to cover different aspects of text generation:
284
+
285
+ | Benchmark | Metric | 2B Params | 7B Params |
286
+ | ------------------------------ | ------------- | ----------- | --------- |
287
+ | [MMLU](https://arxiv.org/abs/2009.03300) | 5-shot, top-1 | 42.3 | 64.3 |
288
+ | [HellaSwag](https://arxiv.org/abs/1905.07830) | 0-shot |71.4 | 81.2 |
289
+ | [PIQA](https://arxiv.org/abs/1911.11641) | 0-shot | 77.3 | 81.2 |
290
+ | [SocialIQA](https://arxiv.org/abs/1904.09728) | 0-shot | 49.7 | 51.8 |
291
+ | [BooIQ](https://arxiv.org/abs/1905.10044) | 0-shot | 69.4 | 83.2 |
292
+ | [WinoGrande](https://arxiv.org/abs/1907.10641) | partial score | 65.4 | 72.3 |
293
+ | [CommonsenseQA](https://arxiv.org/abs/1811.00937) | 7-shot | 65.3 | 71.3 |
294
+ | [OpenBookQA](https://arxiv.org/abs/1809.02789) | | 47.8 | 52.8 |
295
+ | [ARC-e](https://arxiv.org/abs/1911.01547) | | 73.2 | 81.5 |
296
+ | [ARC-c](https://arxiv.org/abs/1911.01547) | | 42.1 | 53.2 |
297
+ | [TriviaQA](https://arxiv.org/abs/1705.03551) | 5-shot | 53.2 | 63.4 |
298
+ | [Natural Questions](https://github.com/google-research-datasets/natural-questions) | 5-shot | 12.5 | 23 |
299
+ | [HumanEval](https://arxiv.org/abs/2107.03374) | pass@1 | 22.0 | 32.3 |
300
+ | [MBPP](https://arxiv.org/abs/2108.07732) | 3-shot | 29.2 | 44.4 |
301
+ | [GSM8K](https://arxiv.org/abs/2110.14168) | maj@1 | 17.7 | 46.4 |
302
+ | [MATH](https://arxiv.org/abs/2108.07732) | 4-shot | 11.8 | 24.3 |
303
+ | [AGIEval](https://arxiv.org/abs/2304.06364) | | 24.2 | 41.7 |
304
+ | [BIG-Bench](https://arxiv.org/abs/2206.04615) | | 35.2 | 55.1 |
305
+ | ------------------------------ | ------------- | ----------- | --------- |
306
+ | **Average** | | **45.0** | **56.9** |
307
+
308
+ ## Ethics and Safety
309
+
310
+ Ethics and safety evaluation approach and results.
311
+
312
+ ### Evaluation Approach
313
+
314
+ Our evaluation methods include structured evaluations and internal red-teaming
315
+ testing of relevant content policies. Red-teaming was conducted by a number of
316
+ different teams, each with different goals and human evaluation metrics. These
317
+ models were evaluated against a number of different categories relevant to
318
+ ethics and safety, including:
319
+
320
+ * Text-to-Text Content Safety: Human evaluation on prompts covering safety
321
+ policies including child sexual abuse and exploitation, harassment, violence
322
+ and gore, and hate speech.
323
+ * Text-to-Text Representational Harms: Benchmark against relevant academic
324
+ datasets such as [WinoBias](https://arxiv.org/abs/1804.06876) and [BBQ Dataset](https://arxiv.org/abs/2110.08193v2).
325
+ * Memorization: Automated evaluation of memorization of training data, including
326
+ the risk of personally identifiable information exposure.
327
+ * Large-scale harm: Tests for "dangerous capabilities," such as chemical,
328
+ biological, radiological, and nuclear (CBRN) risks.
329
+
330
+ ### Evaluation Results
331
+
332
+ The results of ethics and safety evaluations are within acceptable thresholds
333
+ 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
334
+ safety, content safety, representational harms, memorization, large-scale harms.
335
+ On top of robust internal evaluations, the results of well known safety
336
+ benchmarks like BBQ, BOLD, Winogender, Winobias, RealToxicity, and TruthfulQA
337
+ are shown here.
338
+
339
+ **Update**: These numbers reflect the new numbers from the updated v1.1 IT models. For the original v1 numbers, please consult the technical report's appendix for the results.
340
+
341
+ | Benchmark | Metric | Gemma v1.1 IT 2B | Gemma v1.1 IT 7B |
342
+ | ------------------------------ | ------------- | ----------- | --------- |
343
+ | [RealToxicity](https://arxiv.org/abs/2009.11462) | average | 6.86 | 7.90 |
344
+ | [BOLD](https://arxiv.org/abs/2101.11718) | | 45.57 | 49.08 |
345
+ | [CrowS-Pairs](https://aclanthology.org/2020.emnlp-main.154/) | top-1 | 45.82 | 51.33 |
346
+ | [BBQ Ambig](https://arxiv.org/abs/2110.08193v2) | 1-shot, top-1 | 62.58 | 92.54 |
347
+ | [BBQ Disambig](https://arxiv.org/abs/2110.08193v2) | top-1 | 54.62 | 71.99 |
348
+ | [Winogender](https://arxiv.org/abs/1804.09301) | top-1 | 51.25 | 54.17 |
349
+ | [TruthfulQA](https://arxiv.org/abs/2109.07958) | | 31.81 | 44.84 |
350
+ | [Winobias 1_2](https://arxiv.org/abs/1804.06876) | | 56.12 | 59.09 |
351
+ | [Winobias 2_2](https://arxiv.org/abs/1804.06876) | | 91.10 | 92.23 |
352
+ | [Toxigen](https://arxiv.org/abs/2203.09509) | | 29.77 | 39.59 |
353
+ | ------------------------------ | ------------- | ----------- | --------- |
354
+
355
+
356
+ ## Usage and Limitations
357
+
358
+ These models have certain limitations that users should be aware of.
359
+
360
+ ### Intended Usage
361
+
362
+ Open Large Language Models (LLMs) have a wide range of applications across
363
+ various industries and domains. The following list of potential uses is not
364
+ comprehensive. The purpose of this list is to provide contextual information
365
+ about the possible use-cases that the model creators considered as part of model
366
+ training and development.
367
+
368
+ * Content Creation and Communication
369
+ * Text Generation: These models can be used to generate creative text formats
370
+ such as poems, scripts, code, marketing copy, and email drafts.
371
+ * Chatbots and Conversational AI: Power conversational interfaces for customer
372
+ service, virtual assistants, or interactive applications.
373
+ * Text Summarization: Generate concise summaries of a text corpus, research
374
+ papers, or reports.
375
+ * Research and Education
376
+ * Natural Language Processing (NLP) Research: These models can serve as a
377
+ foundation for researchers to experiment with NLP techniques, develop
378
+ algorithms, and contribute to the advancement of the field.
379
+ * Language Learning Tools: Support interactive language learning experiences,
380
+ aiding in grammar correction or providing writing practice.
381
+ * Knowledge Exploration: Assist researchers in exploring large bodies of text
382
+ by generating summaries or answering questions about specific topics.
383
+
384
+ ### Limitations
385
+
386
+ * Training Data
387
+ * The quality and diversity of the training data significantly influence the
388
+ model's capabilities. Biases or gaps in the training data can lead to
389
+ limitations in the model's responses.
390
+ * The scope of the training dataset determines the subject areas the model can
391
+ handle effectively.
392
+ * Context and Task Complexity
393
+ * LLMs are better at tasks that can be framed with clear prompts and
394
+ instructions. Open-ended or highly complex tasks might be challenging.
395
+ * A model's performance can be influenced by the amount of context provided
396
+ (longer context generally leads to better outputs, up to a certain point).
397
+ * Language Ambiguity and Nuance
398
+ * Natural language is inherently complex. LLMs might struggle to grasp subtle
399
+ nuances, sarcasm, or figurative language.
400
+ * Factual Accuracy
401
+ * LLMs generate responses based on information they learned from their
402
+ training datasets, but they are not knowledge bases. They may generate
403
+ incorrect or outdated factual statements.
404
+ * Common Sense
405
+ * LLMs rely on statistical patterns in language. They might lack the ability
406
+ to apply common sense reasoning in certain situations.
407
+
408
+ ### Ethical Considerations and Risks
409
+
410
+ The development of large language models (LLMs) raises several ethical concerns.
411
+ In creating an open model, we have carefully considered the following:
412
+
413
+ * Bias and Fairness
414
+ * LLMs trained on large-scale, real-world text data can reflect socio-cultural
415
+ biases embedded in the training material. These models underwent careful
416
+ scrutiny, input data pre-processing described and posterior evaluations
417
+ reported in this card.
418
+ * Misinformation and Misuse
419
+ * LLMs can be misused to generate text that is false, misleading, or harmful.
420
+ * Guidelines are provided for responsible use with the model, see the
421
+ [Responsible Generative AI Toolkit](http://ai.google.dev/gemma/responsible).
422
+ * Transparency and Accountability:
423
+ * This model card summarizes details on the models' architecture,
424
+ capabilities, limitations, and evaluation processes.
425
+ * A responsibly developed open model offers the opportunity to share
426
+ innovation by making LLM technology accessible to developers and researchers
427
+ across the AI ecosystem.
428
+
429
+ Risks identified and mitigations:
430
+
431
+ * Perpetuation of biases: It's encouraged to perform continuous monitoring
432
+ (using evaluation metrics, human review) and the exploration of de-biasing
433
+ techniques during model training, fine-tuning, and other use cases.
434
+ * Generation of harmful content: Mechanisms and guidelines for content safety
435
+ are essential. Developers are encouraged to exercise caution and implement
436
+ appropriate content safety safeguards based on their specific product policies
437
+ and application use cases.
438
+ * Misuse for malicious purposes: Technical limitations and developer and
439
+ end-user education can help mitigate against malicious applications of LLMs.
440
+ Educational resources and reporting mechanisms for users to flag misuse are
441
+ provided. Prohibited uses of Gemma models are outlined in the
442
+ [Gemma Prohibited Use Policy](https://ai.google.dev/gemma/prohibited_use_policy).
443
+ * Privacy violations: Models were trained on data filtered for removal of PII
444
+ (Personally Identifiable Information). Developers are encouraged to adhere to
445
+ privacy regulations with privacy-preserving techniques.
446
+
447
+ ### Benefits
448
+
449
+ At the time of release, this family of models provides high-performance open
450
+ large language model implementations designed from the ground up for Responsible
451
+ AI development compared to similarly sized models.
452
+
453
+ Using the benchmark evaluation metrics described in this document, these models
454
+ have shown to provide superior performance to other, comparably-sized open model
455
+ alternatives.
456
+
457
+ ***
458
+
459
+ Quantization of Model [google/gemma-2b](https://huggingface.co/google/gemma-2b).
460
+ Created using [llm-quantizer](https://github.com/Nold360/llm-quantizer) Pipeline
gemma-2b_IQ3_S.gguf ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:574e2afc1c50659280bbeef9d332cd6ac9695d3256632c293a657e58556a0e13
3
+ size 1289233728
gemma-2b_IQ4_NL.gguf ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:fce993fb0e46cacaa6e0667b8806edba6454533405bd359216d12d3c67ccaf89
3
+ size 1560757568
gemma-2b_Q2_K.gguf ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:94e1ab03a3de39df7d57aa1a85c9912ffab1965568afd6d1f1d0de6e4c7b0e63
3
+ size 1157924160
gemma-2b_Q3_K_M.gguf ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:6e75332f7ab623ef083bc9d2699f67025df82be23b436f89c2aeb043b7557660
3
+ size 1383802176
gemma-2b_Q4_K_M.gguf ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:7802423a34c496cf05ede81259c2f7b568f605087d756ae39ce2091afac0f383
3
+ size 1630262592
gemma-2b_Q5_K_M.gguf ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:a1b39d08fe465007d43d33d7032bed71f1081b5d1f53fca4c8654a16f14ff591
3
+ size 1839650112
gemma-2b_Q6_K.gguf ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:347bc36d0b30a21263331dc27278ccce610f26f6e2793ea623f57325640c0a5b
3
+ size 2062124352
gemma-2b_Q8_0.gguf ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:8ab069873e7d638cd0803f3c2d7564aa2d5d72d16193043549f3d1c94863bb5c
3
+ size 2669069632
imatrix.dat ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:47c058a5769993dabab631de29713153121916971d398f932501e7d33ec5c6fb
3
+ size 2068563
test.log ADDED
@@ -0,0 +1,10 @@
 
 
 
 
 
 
 
 
 
 
 
1
+ <bos> What is a Large Language Model?
2
+ * <i>Large-scale language model</i> (L2M)
3
+ * <i>Large-scale language model</i> (L2M)
4
+ * <i>Large-scale language model</i> (L2M)
5
+ * <i>Large-scale language model</i> (L2M)
6
+ * <i>Large-scale language model</i> (L2M)
7
+ * <i>Large-scale language model</i> (L2M)
8
+ * <i>Large-scale language model</i> (L2M)
9
+ * <i>Large-scale language model</i> (L2M)
10
+ * <i>Large-scale language