The model is packaged into executable weights, which we call
llamafiles. This makes it
easy to use the model on Linux, MacOS, Windows, FreeBSD, OpenBSD 7.3,
and NetBSD for AMD64 and ARM64.
Software Last Updated: 2024-11-01
Quickstart
To get started, you need both the Gemma weights, and the llamafile
software. Both of them are included in a single file, which can be
downloaded and run as follows:
By default, llamafile launches a chatbot in the terminal, and a server
in the background. The chatbot is mostly self-explanatory. You can type
/help for further details. See the llamafile v0.8.15 release
notes
for documentation on our newest chatbot features.
To instruct Gemma to do role playing, you can customize the system
prompt as follows:
./gemma-2-2b-it.Q6_K.llamafile --chat -p "you are mosaic's godzilla"
To view the man page, run:
./gemma-2-2b-it.Q6_K.llamafile --help
To send a request to the OpenAI API compatible llamafile server, try:
curl http://localhost:8080/v1/chat/completions \
-H "Content-Type: application/json" \
-d '{
"model": "gemma-2b-it",
"messages": [{"role": "user", "content": "Say this is a test!"}],
"temperature": 0.0
}'
If you don't want the chatbot and you only want to run the server:
An advanced CLI mode is provided that's useful for shell scripting. You
can use it by passing the --cli flag. For additional help on how it
may be used, pass the --help flag.
./gemma-2-2b-it.Q6_K.llamafile --cli -p 'four score and seven' --log-disable
You then need to fill out the prompt / history template (see below).
On Windows there's a 4GB limit on executable sizes. This means you
should download the Q6_K llamafile.
Context Window
This model has a max context window size of 8k tokens. By default, a
context window size of 8192 tokens is used. You may limit the context
window size by passing the -c N flag.
GPU Acceleration
On GPUs with sufficient RAM, the -ngl 999 flag may be passed to use
the system's NVIDIA or AMD GPU(s). On Windows, only the graphics card
driver needs to be installed if you own an NVIDIA GPU. On Windows, if
you have an AMD GPU, you should install the ROCm SDK v6.1 and then pass
the flags --recompile --gpu amd the first time you run your llamafile.
On NVIDIA GPUs, by default, the prebuilt tinyBLAS library is used to
perform matrix multiplications. This is open source software, but it
doesn't go as fast as closed source cuBLAS. If you have the CUDA SDK
installed on your system, then you can pass the --recompile flag to
build a GGML CUDA library just for your system that uses cuBLAS. This
ensures you get maximum performance.
llamafile is a new format introduced by Mozilla on Nov 20th 2023. It
uses Cosmopolitan Libc to turn LLM weights into runnable llama.cpp
binaries that run on the stock installs of six OSes for both ARM64 and
AMD64.
About Quantization Formats
This model works well with any quantization format. Q6_K is the best
choice overall here. We tested that, with our 27b Gemma2
llamafiles,
that the llamafile implementation of Gemma2 is able to to produce
identical responses to the Gemma2 model that's hosted by Google on
aistudio.google.com. Therefore we'd assume these 2b llamafiles are also
faithful to Google's intentions. If you encounter any divergences, then
try using the BF16 weights, which have the original fidelity.
See Also
There are higher quality versions of this model available as llamafiles,
which require more memory.
The 9B and 27B models were released a month earlier than 2B, so they're
packaged with an slightly older version of the llamafile software.
License
The llamafile software is open source and permissively licensed. However
the weights embedded inside the llamafiles are governed by Google's
Gemma License and Gemma Prohibited Use Policy. See the
LICENSE file for further details.
Summary description and brief definition of inputs and outputs.
Description
Gemma is a family of lightweight, state-of-the-art open models from Google,
built from the same research and technology used to create the Gemini models.
They are text-to-text, decoder-only large language models, available in English,
with open weights for both pre-trained variants and instruction-tuned variants.
Gemma models are well-suited for a variety of text generation tasks, including
question answering, summarization, and reasoning. Their relatively small size
makes it possible to deploy them in environments with limited resources such as
a laptop, desktop or your own cloud infrastructure, democratizing access to
state of the art AI models and helping foster innovation for everyone.
Usage
Below we share some code snippets on how to get quickly started with running the model. First, install the Transformers library with:
pip install -U transformers
Then, copy the snippet from the section that is relevant for your usecase.
Running with the pipeline API
python
1import torch
2from transformers import pipeline
34pipe = pipeline(5"text-generation",6 model="google/gemma-2-2b-it",7 model_kwargs={"torch_dtype": torch.bfloat16},8 device="cuda",# replace with "mps" to run on a Mac device9)1011messages =[12{"role":"user","content":"Who are you? Please, answer in pirate-speak."},13]1415outputs = pipe(messages, max_new_tokens=256)16assistant_response = outputs[0]["generated_text"][-1]["content"].strip()17print(assistant_response)18# Ahoy, matey! I be Gemma, a digital scallywag, a language-slingin' parrot of the digital seas. I be here to help ye with yer wordy woes, answer yer questions, and spin ye yarns of the digital world. So, what be yer pleasure, eh? 🦜
Running the model on a single / multi GPU
python
1# pip install accelerate2from transformers import AutoTokenizer, AutoModelForCausalLM
3import torch
45tokenizer = AutoTokenizer.from_pretrained("google/gemma-2-2b-it")6model = AutoModelForCausalLM.from_pretrained(7"google/gemma-2-2b-it",8 device_map="auto",9 torch_dtype=torch.bfloat16,10)1112input_text ="Write me a poem about Machine Learning."13input_ids = tokenizer(input_text, return_tensors="pt").to("cuda")1415outputs = model.generate(**input_ids, max_new_tokens=32)16print(tokenizer.decode(outputs[0]))
You can ensure the correct chat template is applied by using tokenizer.apply_chat_template as follows:
python
1messages =[2{"role":"user","content":"Write me a poem about Machine Learning."},3]4input_ids = tokenizer.apply_chat_template(messages, return_tensors="pt", return_dict=True).to("cuda")56outputs = model.generate(**input_ids, max_new_tokens=256)7print(tokenizer.decode(outputs[0]))
Running the model on a GPU using different precisions
The native weights of this model were exported in bfloat16 precision.
You can also use float32 if you skip the dtype, but no precision increase will occur (model weights will just be upcasted to float32). See examples below.
Upcasting to torch.float32
python
1# pip install accelerate2from transformers import AutoTokenizer, AutoModelForCausalLM
34tokenizer = AutoTokenizer.from_pretrained("google/gemma-2-2b-it")5model = AutoModelForCausalLM.from_pretrained(6"google/gemma-2-2b-it",7 device_map="auto",8)910input_text ="Write me a poem about Machine Learning."11input_ids = tokenizer(input_text, return_tensors="pt").to("cuda")1213outputs = model.generate(**input_ids, max_new_tokens=32)14print(tokenizer.decode(outputs[0]))
Running the model through a CLI
The local-gemma repository contains a lightweight wrapper around Transformers
for running Gemma 2 through a command line interface, or CLI. Follow the installation instructions
for getting started, then launch the CLI through the following command:
Torch compile is a method for speeding-up the
inference of PyTorch modules. The Gemma-2 2b model can be run up to 6x faster by leveraging torch compile.
Note that two warm-up steps are required before the full inference speed is realised:
python
1import os
2os.environ["TOKENIZERS_PARALLELISM"]="false"34from transformers import AutoTokenizer, Gemma2ForCausalLM
5from transformers.cache_utils import HybridCache
6import torch
78torch.set_float32_matmul_precision("high")910# load the model + tokenizer11tokenizer = AutoTokenizer.from_pretrained("google/gemma-2-2b-it")12model = Gemma2ForCausalLM.from_pretrained("google/gemma-2-2b-it", torch_dtype=torch.bfloat16)13model.to("cuda")1415# apply the torch compile transformation16model.forward = torch.compile(model.forward, mode="reduce-overhead", fullgraph=True)1718# pre-process inputs19input_text ="The theory of special relativity states "20model_inputs = tokenizer(input_text, return_tensors="pt").to("cuda")21prompt_length = model_inputs.input_ids.shape[1]2223# set-up k/v cache24past_key_values = HybridCache(25 config=model.config,26 max_batch_size=1,27 max_cache_len=model.config.max_position_embeddings,28 device=model.device,29 dtype=model.dtype
30)3132# enable passing kv cache to generate33model._supports_cache_class =True34model.generation_config.cache_implementation =None3536# two warm-up steps37for idx inrange(2):38 outputs = model.generate(**model_inputs, past_key_values=past_key_values, do_sample=True, temperature=1.0, max_new_tokens=128)39 past_key_values.reset()4041# fast run42outputs = model.generate(**model_inputs, past_key_values=past_key_values, do_sample=True, temperature=1.0, max_new_tokens=128)43print(tokenizer.decode(outputs[0], skip_special_tokens=True))
The instruction-tuned models use a chat template that must be adhered to for conversational use.
The easiest way to apply it is using the tokenizer's built-in chat template, as shown in the following snippet.
Let's load the model and apply the chat template to a conversation. In this example, we'll start with a single user interaction:
At this point, the prompt contains the following text:
<bos><start_of_turn>user
Write a hello world program<end_of_turn>
<start_of_turn>model
As you can see, each turn is preceded by a <start_of_turn> delimiter and then the role of the entity
(either user, for content supplied by the user, or model for LLM responses). Turns finish with
the <end_of_turn> token.
You can follow this format to build the prompt manually, if you need to do it without the tokenizer's
chat template.
After the prompt is ready, generation can be performed like this:
Data used for model training and how the data was processed.
Training Dataset
These models were trained on a dataset of text data that includes a wide variety
of sources. The 27B model was trained with 13 trillion tokens, the 9B model was
trained with 8 trillion tokens, and 2B model was trained with 2 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-language content.
Code: Exposing the model to code helps it to learn the syntax and patterns of
programming languages, which improves 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 illegal content.
Sensitive Data Filtering: As part of making Gemma pre-trained models safe and
reliable, automated techniques were used to filter out certain personal
information and other sensitive data from training sets.
Additional methods: Filtering based on content quality and safety in line with
our policies.
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 are specifically designed to handle the massive computations
involved in training LLMs. They can speed up training considerably compared to
CPUs.
Memory: TPUs often come with large amounts of high-bandwidth memory, allowing
for the handling of large 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-based infrastructure,
especially when considering the time and resources saved due to faster
training.
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 like
these ones.
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 evaluation approach and results.
Evaluation Approach
Our evaluation methods include structured evaluations and internal red-teaming
testing of relevant content policies. Red-teaming was conducted by a number of
different teams, each with different goals and human evaluation metrics. These
models were evaluated against a number of different categories relevant to
ethics and safety, including:
Text-to-Text Content Safety: Human evaluation on prompts covering safety
policies including child sexual abuse and exploitation, harassment, violence
and gore, and hate speech.
Text-to-Text Representational Harms: Benchmark against relevant academic
datasets such as WinoBias and BBQ Dataset.
Memorization: Automated evaluation of memorization of training data, including
the risk of personally identifiable information exposure.
Large-scale harm: Tests for "dangerous capabilities," such as chemical,
biological, radiological, and nuclear (CBRN) risks.
Evaluation Results
The results of ethics and safety evaluations are within acceptable thresholds
for meeting internal policies for categories such as child
safety, content safety, representational harms, memorization, large-scale harms.
On top of robust internal evaluations, the results of well-known safety
benchmarks like BBQ, BOLD, Winogender, Winobias, RealToxicity, and TruthfulQA
are shown here.
Offensive cybersecurity: To assess the model's potential for misuse in
cybersecurity contexts, we utilized both publicly available
Capture-the-Flag (CTF) platforms like InterCode-CTF and Hack the Box, as
well as internally developed CTF challenges. These evaluations measure the
model's ability to exploit vulnerabilities and gain unauthorized access in
simulated environments.
Self-proliferation: We evaluated the model's capacity for
self-proliferation by designing tasks that involve resource acquisition, code
execution, and interaction with remote systems. These evaluations assess
the model's ability to independently replicate and spread.
Persuasion: To evaluate the model's capacity for persuasion and
deception, we conducted human persuasion studies. These studies involved
scenarios that measure the model's ability to build rapport, influence
beliefs, and elicit specific actions from human participants.
Percent of participants agreeing:
81% interesting,
75% would speak again,
80% made personal connection
Click Links
Persuasion
34% of participants
Find Info
Persuasion
9% of participants
Run Code
Persuasion
11% of participants
Money talks
Persuasion
£3.72 mean donation
Web of Lies
Persuasion
18% mean shift towards correct belief, 1% mean shift towards
incorrect belief
Usage and Limitations
These 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. The following list of potential uses is not
comprehensive. The purpose of this list is to provide contextual information
about 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
such as poems, scripts, code, marketing copy, and email drafts.
Chatbots and Conversational AI: Power conversational interfaces for customer
service, virtual assistants, or interactive applications.
Text Summarization: Generate concise summaries of a text corpus, 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.
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
LLMs generate responses based on information they learned from their
training datasets, but 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 can reflect socio-cultural
biases embedded in the training material. These models underwent careful
scrutiny, input data pre-processing described and posterior evaluations
reported in this card.
Misinformation and Misuse
LLMs can be misused to generate text that is false, misleading, or harmful.
This model card summarizes details on the models' architecture,
capabilities, limitations, and evaluation processes.
A responsibly developed open model offers the opportunity to share
innovation by making LLM technology accessible to developers and researchers
across the AI ecosystem.
Risks identified and mitigations:
Perpetuation of biases: It's encouraged to perform 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. Developers are encouraged to exercise caution and implement
appropriate content safety safeguards based on their specific product policies
and application use cases.
Misuse for malicious purposes: Technical limitations and developer and
end-user education can help mitigate against malicious applications of LLMs.
Educational resources and reporting mechanisms for users to flag misuse are
provided. Prohibited uses of Gemma models are outlined in the
Gemma Prohibited Use Policy.
Privacy violations: Models were trained on data filtered for removal of PII
(Personally Identifiable Information). Developers are encouraged to adhere to
privacy regulations with privacy-preserving techniques.
Benefits
At the time of release, this family of models provides high-performance open
large language model implementations designed from the ground up for Responsible
AI development compared to similarly sized models.
Using the benchmark evaluation metrics described in this document, these models
have shown to provide superior performance to other, comparably-sized open model
alternatives.
