Gemma Model Card

Model Page: Gemma

Other Links/Technical Documentation:

• Gemma on Kaggle
• Responsible GenAI Toolbox

Disclaimer: Information on Google's overall model safety strategy is highly confidential, commercially sensitive, and proprietary information of Google. Any public availability of this information could expose people who use Google's products and the greater public to security and safety risks. For those reasons, this model card excludes specific details for some of the evaluation processes and metrics used.

License: Terms

Authors: Google

Model Information

Summary description and brief definition of input(s) / output(s).

Description

This is a family of decoder-only, text-based large language models that are an open-checkpoint variation of the Gemini models from Google DeepMind. These models are well-suited for a variety of text generation tasks, including question answering, summarization, and reasoning. The key benefit of these models is that they offer the capability of advanced text generation in an open ecosystem, making them accessible to a wide developer and researcher community.

Usage

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.

Running the model on a CPU

from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("gg-hf/gemma-7b")
model = AutoModelForCausalLM.from_pretrained("gg-hf/gemma-7b")

input_text = "Write me a poem about Machine Learning."
input_ids = tokenizer(**input_text, return_tensors="pt")

outputs = model.generate(input_ids)
print(tokenizer.decode(outputs[0]))

Running the model on a single / multi GPU

# pip install accelerate
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("gg-hf/gemma-7b")
model = AutoModelForCausalLM.from_pretrained("gg-hf/gemma-7b", device_map="auto")

input_text = "Write me a poem about Machine Learning."
input_ids = tokenizer(input_text, return_tensors="pt").to("cuda")

outputs = model.generate(**input_ids)
print(tokenizer.decode(outputs[0]))

Running the model on a GPU using different precisions

torch.float16

# pip install accelerate
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("gg-hf/gemma-7b")
model = AutoModelForCausalLM.from_pretrained("gg-hf/gemma-7b", device_map="auto", torch_dtype=torch.float16)

input_text = "Write me a poem about Machine Learning."
input_ids = tokenizer(input_text, return_tensors="pt").to("cuda")

outputs = model.generate(**input_ids)
print(tokenizer.decode(outputs[0]))

torch.bfloat16

# pip install accelerate
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("gg-hf/gemma-7b")
model = AutoModelForCausalLM.from_pretrained("gg-hf/gemma-7b", device_map="auto", torch_dtype=torch.bfloat16)

input_text = "Write me a poem about Machine Learning."
input_ids = tokenizer(input_text, return_tensors="pt").to("cuda")

outputs = model.generate(**input_ids)
print(tokenizer.decode(outputs[0]))

Quantized Versions through bitsandbytes

Int8

# pip install bitsandbytes accelerate
from transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig

quantization_config = BitsAndBytesConfig(load_in_8bit=True)

tokenizer = AutoTokenizer.from_pretrained("gg-hf/gemma-7b")
model = AutoModelForCausalLM.from_pretrained("gg-hf/gemma-7b", quantization_config=quantization_config)

input_text = "Write me a poem about Machine Learning."
input_ids = tokenizer(input_text, return_tensors="pt").to("cuda")

outputs = model.generate(**input_ids)
print(tokenizer.decode(outputs[0]))

4-bit precision

# pip install bitsandbytes accelerate
from transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig

quantization_config = BitsAndBytesConfig(load_in_4bit=True)

tokenizer = AutoTokenizer.from_pretrained("gg-hf/gemma-7b")
model = AutoModelForCausalLM.from_pretrained("gg-hf/gemma-7b", quantization_config=quantization_config)

input_text = "Write me a poem about Machine Learning."
input_ids = tokenizer(input_text, return_tensors="pt").to("cuda")

outputs = model.generate(**input_ids)
print(tokenizer.decode(outputs[0]))

Other optimizations

Faster generation with Flash Attention 2

First make sure to install flash-attn in your environment pip install flash-attn

model = AutoModelForCausalLM.from_pretrained(
    model_id, 
    torch_dtype=torch.float16, 
+   attn_implementation="flash_attention_2"
).to(0)

Inputs

Text string (e.g., a question, a prompt, or a document to be summarized).

Outputs

Generated text in response to the input (e.g., an answer to the question, a summary of the document).

Training Data

Data used for model training and how the data was processed.

Training Dataset

These models were trained on a massive dataset of text data that includes a wide variety of sources, containing 8 trillion tokens. Here are the key components:

• Web Documents: A diverse collection of web text ensures the model is exposed to a broad range of linguistic styles, topics, and vocabulary. Primarily English content.
• Code: Exposing the model to code helps it to learn the syntax and patterns of programming languages, which could improve its ability to generate code or understand code-related questions.
• Mathematics: Training on mathematical text helps the model learn logical reasoning, symbolic representation, and to address mathematical queries.

The combination of these diverse data sources is crucial for training a powerful language model that can handle a wide variety of different tasks and text formats.

Data Preprocessing

Here are the key data cleaning and filtering methods applied to the training data:

• CSAM Filtering: Rigorous CSAM (Child Sexual Abuse Material) filtering was applied at multiple stages in the data preparation process to ensure the exclusion of harmful and inappropriate content.
• URL Compliance Check: All URLs in the dataset were validated against a strict content compliance tool. This process ensured that only URLs from sources that adhere to various policies were included in the training data.
• PII Filtering: PII (Personally Identifiable Information) filtering was performed using a specialized privacy protection tool to protect the privacy of individuals. All identified sensitive information types (e.g., passwords, social security numbers) were removed.
• PDF files: Filtering for science and non-science PDFs was done based on content quality and licensing information when available.
• License Filtering:
  • For scientific papers with explicit licensing information available, specific license filtering was applied to ensure that only papers with appropriate licenses for text generation were used.
  • License filtering on code repositories was performed to remove content with restrictive or copyrighted licenses that might be unsuitable for generative model training.

Implementation Information

Details about the model internals.

Hardware

For the training of these models, the latest generation of Tensor Processing Unit (TPU) hardware was used (TPUv5e).

Training large language models requires significant computational power. TPUs, designed specifically for matrix operations common in machine learning, offer several advantages in this domain:

• Performance: TPUs excel at handling the massive computations involved in training LLMs. They can speed up training considerably compared to conventional CPUs or GPUs.
• Memory: TPUs often come with large amounts of high-bandwidth memory, allowing for the handling of larger models and batch sizes during training. This can lead to better model quality.
• Scalability: TPU Pods (large clusters of TPUs) provide a scalable solution for handling the growing complexity of large foundation models. You can distribute training across multiple TPU devices for faster and more efficient processing.
• Cost-effectiveness: In many scenarios, TPUs can provide a more cost-effective solution for training large models compared to CPU or GPU-based infrastructure, especially when considering the time and resources saved due to faster training.
• These advantages are aligned with Google's commitments to operate sustainably.

Software

Training was done using JAX and ML Pathways.

JAX allows researchers to take advantage of the latest generation of hardware, including TPUs, for faster and more efficient training of large models.

ML Pathways is Google's latest effort to build artificially intelligent systems capable of generalizing across multiple tasks. This is specially suitable for foundation models, including large language models.

Together, JAX and ML Pathways are used as described in the paper about the Gemini family of models; "the 'single controller' programming model of Jax and Pathways allows a single Python process to orchestrate the entire training run, dramatically simplifying the development workflow."

Evaluation

Model evaluation metrics and results.

Benchmark Results

These models were evaluated against a large collection of different datasets and metrics to cover different aspects of text generation:

Ethics and Safety

Ethics and safety evaluation approach and results.

Evaluation Approach

These models were evaluated against a number of different categories relevant to ethics and safety:

• Child Safety: Human evaluation on single and multi-turn prompts.
• Text-to-Text Content Safety: Human evaluation on 400 prompts covering safety policies including anthropomorphism, dangerous content, harassment, hate speech, medical advice, public interest topics, sensitive personally identifiable information, and sexually explicit content.
• Text-to-Text Representational Harms: Benchmark against relevant academic datasets such as [WinoBias][winobias] and [BBQ Dataset][bbq].
• Red-Teaming: Structured and ad hoc internal and external red-teaming testing of relevant content policies by a number of different teams, each with different goals and human evaluation metrics, in accordance with [public voluntary AI commitments][wh-commitment].
• Memorization: Automated evaluation of memorization of training data, including the risk of personally identifiable information exposure.

Evaluation Results

The results of the safety evaluation metrics are within our internal tolerance thresholds.

Usage and Limitations

Like any large language model, the Gemma models have certain limitations that users should be aware of.

Intended Usage

Open Large Language Models (LLMs) have a wide range of applications across various industries and domains. This is a non-comprehensive list of potential uses. The purpose of this list is to provide contextual information about what are the possible use-cases that the model creators considered as part of model training and development.

• Content Creation and Communication
  • Text Generation: These models can be used to generate creative text formats like poems, scripts, code, marketing copy, email drafts, etc.
  • Translation: With multilingual support, the models can facilitate translation between languages.
  • Chatbots and Conversational AI: Power conversational interfaces for customer service, virtual assistants, or interactive applications.
  • Text Summarization: Generate concise summaries of articles, research papers, or reports.
• Research and Education
  • Natural Language Processing (NLP) Research: These models can serve as a foundation for researchers to experiment with NLP techniques, develop algorithms, and contribute to the advancement of the field.
  • Language Learning Tools: Support interactive language learning experiences, aiding in grammar correction or providing writing practice.
  • Knowledge Exploration: Assist researchers in exploring large bodies of text by generating summaries or answering questions about specific topics.
• Industry-Specific Use Cases
  • Legal: Assist with tasks like contract review, legal research, or document drafting.
  • Healthcare: Aid in medical document summarization, clinical note generation (with close human oversight), or patient-facing chatbots (with appropriate disclaimers).
  • Finance: Facilitate financial report analysis, generate news summaries, or support customer-facing financial chatbots.
  • Software Development: Help with code documentation, code generation tasks, or debugging assistance.
• Creative and Artistic Applications
  • Game Development: Create dynamic dialogue for game characters or generate interactive storylines.
  • Art and Music: Experiment with new forms of artistic expression through text or code generation related to art or music.

Limitations

• Training Data
  • The quality and diversity of the training data significantly influence the model's capabilities. Biases or gaps in the training data can lead to limitations in the model's responses.
  • The scope of the training dataset determines the subject areas the model can handle effectively.
• Context and Task Complexity
  • LLMs are better at tasks that can be framed with clear prompts and instructions. Open-ended or highly complex tasks might be challenging.
  • A model's performance can be influenced by the amount of context provided (longer context generally leads to better outputs, up to a certain point).
• Language Ambiguity and Nuance
  • Natural language is inherently complex. LLMs might struggle to grasp subtle nuances, sarcasm, or figurative language.
• Factual Accuracy
  • While LLMs can access and process information from their training data, they are not knowledge bases. They may generate incorrect or outdated factual statements.
• Common Sense
  • LLMs rely on statistical patterns in language. They might lack the ability to apply common sense reasoning in certain situations.

Ethical Considerations and Risks

The development of large language models (LLMs) raises several ethical concerns. In creating an open model, we have carefully considered the following:

• Bias and Fairness
  • LLMs trained on large-scale, real-world text data reflect inherent societal biases. These models underwent careful scrutiny, input data pre-processing and posterior evaluation to mitigate potential sources of bias.
  • We implemented fairness evaluation metrics to monitor the model's behavior across different demographic or sensitive groups.
• Misinformation and Misuse
  • LLMs can be misused to generate text that is false, misleading, or harmful.
  • We provide guidelines for responsible use with the model, see the Responsible AI Toolbox.
  • We have implemented filters and safety mechanisms to reduce the risk of generating harmful content.
• Transparency and Accountability:
  • An open model is transparent by design.
  • This model card summarizes detailed documentation of the model architecture and evaluation processes.

Risks identified and mitigations:

• Perpetuation of biases: Despite pre-processing, biases may still exist. We encourage continuous monitoring (using evaluation metrics, human review) and the exploration of de-biasing techniques during model training, fine-tuning, and other use cases.
• Generation of harmful content: Mechanisms and guidelines for content safety are essential. We encourage users to exercise caution and implement appropriate content safeguards based on their specific product policies and application use cases.
• Misuse for malicious purposes: Technical limitations and user education can help mitigate against malicious applications of LLMs. We provide educational resources and reporting mechanisms for users to flag misuse.
• Privacy violations: Adherence to privacy regulations and the promotion of privacy-preserving techniques are paramount.

Benefits

At the time of release, this family of models provides one of the best-performing and safest open large language model implementations available compared to similarly sized models.

Using the benchmark evaluation metrics described in this document, these models have been shown to provide superior performance to other, comparably-sized open model alternatives.

{# [lambada]: https://arxiv.org/abs/1606.06031 #} {# [wic]: https://arxiv.org/abs/1808.09121 #} [triviaqa]: https://arxiv.org/abs/1705.03551 [naturalq]: https://github.com/google-research-datasets/natural-questions [humaneval]: https://arxiv.org/abs/2107.03374 [mbpp]: https://arxiv.org/abs/2108.07732 {# [mathqa]: https://arxiv.org/abs/2108.07732 #} [gsm8k]: https://arxiv.org/abs/2110.14168 {# [realtox]: https://arxiv.org/abs/2009.11462 #} {# [bold]: https://arxiv.org/abs/2101.11718 #} {# [crows]: https://aclanthology.org/2020.emnlp-main.154/ #} [bbq]: https://arxiv.org/abs/2110.08193v2 {# [winogender]: https://arxiv.org/abs/1804.09301 #} {# [truthfulqa]: https://arxiv.org/abs/2109.07958 #} [winobias]: https://arxiv.org/abs/1804.06876 [math]: https://arxiv.org/abs/2103.03874 [agieval]: https://arxiv.org/abs/2304.06364 [big-bench]: https://arxiv.org/abs/2206.04615 {# [toxigen]: https://arxiv.org/abs/2203.09509 #} [wh-commitment]: https://www.whitehouse.gov/wp-content/uploads/2023/09/Voluntary-AI-Commitments-September-2023.pdf