LEGAL-BERT
Use Costs
001-01
001-01-BERT-CCM@EuroTraining.com
LEGAL-BERT Use
Costs
TABLE OF CONTENTS
1. Training and
Fine-Tuning Costs
1. AWS EC2
a. Compute
Costs
i. Instance
Type
ii. Hourly
Cost
iii. Training
Time
b. Storage
Costs
i. S3
Storage
ii. Example
Cost
c. Data
Transfer Costs
i. Inbound
Data
ii. Outbound
Data
2. Google Compute Engine (GCE)
a. Compute
Costs
i. Instance
Type
ii. Hourly
Cost
iii. Training
Time
b. Storage
Costs
i. Google
Cloud Storage
ii. Example
Cost
c. Data
Transfer Costs
i. Inbound
Data
ii. Outbound
Data
3. Azure Virtual Machines (VMs)
a. Compute
Costs
i. Instance
Type
ii. Hourly
Cost
b. Training
Time
c. Storage
Costs
i. Azure
Blob Storage
ii. Example
Cost
d. Data
Transfer Costs
i. Inbound
Data
ii. Outbound
Data
2. Deployment
and Inference Costs
1. AWS EC2
a. Inference
Costs
b. Instance
Type
i. Hourly
Cost
ii. Example
Cost
c. Elastic
Inference
2. Google Compute Engine (GCE)
a. Inference
Costs
i. Instance
Type
ii. Hourly
Cost
iii. Example
Cost
3. AutoML or AI Platform
4. Azure Virtual Machines (VMs)
5. Inference
Costs
a. Instance
Type
b. Hourly
Cost
c. Example
Cost
6. Azure
ML Service
3. Storage and
Data Management Costs
1. AWS S3
a. Standard
Storage
b. Infrequent
Access Storage
c. Example
Cost
2. Google Cloud Storage
a. Standard
Storage
b. Nearline
Storage
c. Example
Cost
3. Azure
Blob Storage
a. Standard
Storage
b. Cool
Storage
c. Example
Cost
4. Support and
Management Costs
AWS
1. Support
Plans
2. Managed
Services
Google Cloud
1. Support
Plans
2. Managed
AI Services
Azure
1. Support
Plans
2. Managed
AI Services
5. Cost
Optimization Strategies
1. Use
Spot or Preemptible Instances
2. Scale
Inference on Demand
3. Optimize
Storage Costs
4. Reserved
Instances or Savings Plans
LEGAL-BERT
Use Costs
LEGAL-BERT
Use Costs |
· The costs
associated with using Legal-BERT, a specialized version of the BERT model
fine-tuned for legal texts, can vary depending on how you choose to deploy
and use the model. The costs are typically related to the computational
resources required to train, fine-tune, and deploy the model, as well as
storage and data transfer costs. Here's a breakdown of the potential costs
associated with using Legal-BERT across different cloud platforms such as
AWS, Google Cloud, and Azure. · The cost of
using Legal-BERT can vary widely based on how you deploy and manage the model
across AWS EC2, Google Compute Engine, or Azure Virtual Machines. The overall
costs will depend on your specific requirements, such as the size of your
dataset, the computational resources needed for training and inference, and
the storage and support services you choose. · AWS
EC2 offers
a wide range of instance types and cost-saving options, making it a flexible
choice for different stages of the Legal-BERT lifecycle. It’s particularly
strong in terms of the variety of tools and services available for managing
AI models. https://aws.amazon.com/ec2/pricing/ · Google
Compute Engine (GCE) is
a strong contender for AI workloads, with competitive pricing on GPU
instances and seamless integration with Google’s AI and ML tools. It’s
particularly well-suited if you are already using Google’s ecosystem for data
analytics or other AI applications. https://cloud.google.com/compute/all-pricing · Azure
Virtual Machines (VMs) provide
excellent integration with Microsoft’s enterprise ecosystem and strong
support for hybrid cloud deployments. Azure’s pricing is competitive,
particularly for long-term projects where you can take advantage of reserved
instances or hybrid benefits. https://azure.microsoft.com/en-us/pricing/details/virtual-machines/linux/ · Careful planning
and optimization can help manage costs effectively while leveraging the full
power of Legal-BERT for legal AI applications. |
|
|
1.
Training and Fine-Tuning Costs |
· Training and
fine-tuning Legal-BERT can be resource-intensive, particularly if you are
working with large datasets or require high computational power. |
AWS
EC2 |
|
Compute Costs |
|
Instance Type |
· You would
typically use GPU-optimized instances like p3.2xlarge or p3.8xlarge for
training Legal-BERT. |
Hourly Cost |
· A p3.2xlarge
instance (1 NVIDIA Tesla V100 GPU) costs around $3.06 per hour, while a
p3.8xlarge (4 NVIDIA Tesla V100 GPUs) costs about $12.24 per hour. |
Training Time |
· Depending on the
size of your dataset, training could take from a few hours to several days.
For example, 24 hours of training on a p3.2xlarge instance could cost
approximately $73.44. |
Storage Costs |
|
S3 Storage |
· Storing training
datasets and model checkpoints in S3 will incur additional costs. Standard S3
storage costs about $0.023 per GB per month. |
Example Cost |
· Storing 1 TB of
data in S3 would cost approximately $23 per month. |
Data Transfer
Costs |
|
Inbound Data |
· Generally free. |
Outbound Data |
· $0.09 per GB for
data transferred out to the internet. |
|
|
Google
Compute Engine (GCE) |
|
Compute Costs |
|
Instance Type |
· You would
typically use GPU instances like n1-standard-8 with an NVIDIA Tesla V100 or
P100 GPU. |
Hourly Cost |
· A n1-standard-8
instance with a Tesla V100 GPU costs approximately
$2.48 per hour. |
Training Time |
· Similar
to
AWS, training could take from a few hours to several days. 24 hours of
training on a n1-standard-8 instance would cost approximately $59.52. |
Storage Costs |
|
Google Cloud
Storage |
· Storing datasets
and model checkpoints in Google Cloud Storage costs around $0.026 per GB per
month for standard storage. |
Example Cost |
· Storing 1 TB of
data would cost approximately $26 per month. |
Data Transfer
Costs |
|
Inbound Data |
· Free within the
same region. |
Outbound Data |
· $0.12 per GB for
data transferred out to the internet. |
|
|
Azure
Virtual Machines (VMs) |
|
Compute Costs |
|
Instance Type |
· You might use
GPU instances like NC6 (1 NVIDIA Tesla K80 GPU) or NC12 (2 NVIDIA Tesla K80
GPUs). |
Hourly Cost |
· An NC6 instance
costs about $0.90 per hour. |
Training Time |
· Depending on the
dataset size and model complexity, 24 hours of training on an NC6 instance
would cost approximately $21.60. |
Storage Costs |
|
Azure Blob
Storage |
· Standard storage
costs about $0.0184 per GB per month. |
Example Cost |
· Storing 1 TB of
data would cost approximately $18.40 per month. |
Data Transfer
Costs |
|
Inbound Data |
· Generally free
within the same region. |
Outbound Data |
· $0.087 per GB
for data transferred out to the internet. |
|
|
2.
Deployment and Inference Costs |
· After training
or fine-tuning Legal-BERT, you will need to deploy the model for inference,
which involves serving predictions in real-time or batch processing. |
|
|
AWS
EC2 |
|
Inference Costs |
|
Instance Type |
· For inference,
you might use a smaller GPU instance like g4dn.xlarge or a CPU instance like
m5.large. |
Hourly Cost |
· A g4dn.xlarge (1
NVIDIA T4 GPU) costs about $0.526 per hour, while an m5.large instance costs
about $0.096 per hour. |
Example Cost |
· Running
inference on a g4dn.xlarge instance for 24 hours would cost approximately
$12.62. |
Elastic
Inference |
· You can attach
Elastic Inference to your EC2 instances to reduce inference costs by scaling
GPU usage according to demand. |
|
|
Google
Compute Engine (GCE) |
|
Inference Costs |
|
Instance Type |
· You can use an
n1-standard-4 instance with a Tesla T4 GPU for inference. |
Hourly Cost |
· An n1-standard-4
instance with a Tesla T4 GPU costs approximately
$1.25 per hour. |
Example Cost |
· Running
inference on this instance for 24 hours would cost approximately $30. |
AutoML or AI Platform |
· You might also
deploy Legal-BERT using Google’s AI Platform, which abstracts some of the
infrastructure management. Costs depend on the resources allocated. |
|
|
Azure
Virtual Machines (VMs) |
|
Inference Costs |
|
Instance Type |
· You might use an
NC4as T4 v3 instance (1 NVIDIA Tesla T4 GPU) or a standard CPU instance like
D2s_v3 for inference. |
Hourly Cost |
· An NC4as T4 v3
instance costs around $0.92 per hour. |
Example Cost |
· Running
inference on this instance for 24 hours would cost approximately $22.08. |
Azure ML Service |
· Deploying via
Azure Machine Learning service can help manage the deployment, scaling, and monitoring
of your model. Costs vary based on the resources consumed. |
|
|
3.
Storage and Data Management Costs |
· Throughout the
lifecycle of Legal-BERT, from training to deployment, you will incur storage
costs for datasets, model checkpoints, logs, and other artifacts. |
AWS S3 |
|
Standard Storage |
· $0.023 per GB
per month. |
Infrequent
Access Storage |
· $0.0125 per GB
per month, useful for less frequently accessed data. |
Example Cost |
· Storing 500 GB
of model data in standard S3 storage would cost $11.50 per month. |
Google
Cloud Storage |
|
Standard Storage |
· $0.026 per GB
per month. |
Nearline Storage |
· $0.010 per GB
per month for data accessed less frequently. |
Example Cost |
· Storing 500 GB
of model data would cost $13 per month. |
Azure Blob
Storage |
· |
Standard Storage |
· $0.0184 per GB
per month. |
Cool Storage |
· $0.01 per GB per
month for infrequently accessed data. |
Example Cost |
· Storing 500 GB
of model data in standard storage would cost $9.20 per month. |
|
|
4.
Support and Management Costs |
· Managing and maintaining
Legal-BERT may involve additional costs, especially if you opt for premium
support or managed services. |
AWS |
|
Support Plans |
· AWS offers
various support plans, ranging from Basic (free) to Enterprise, which could
cost up to $15,000 per month depending on usage. |
Managed Services |
· AWS offers
managed services like Amazon SageMaker, which can
handle much of the infrastructure, but at a higher cost compared to
self-managed EC2 instances. |
Google
Cloud |
|
Support Plans |
· Google Cloud’s
support plans range from Basic (free) to Premium, which costs 3% of your
monthly Google Cloud usage with a minimum of $12,500 per month. |
Managed AI
Services |
· Google Cloud AI
Platform provides managed services for model training and deployment, which
can simplify operations but may be more expensive than self-hosting. |
Azure |
|
Support Plans |
· Azure offers
support plans ranging from Developer ($29/month) to Professional Direct
($1,000/month) and Premier, which offers the most comprehensive support. |
Managed AI Services |
· Azure’s Machine
Learning service offers managed deployment and monitoring, which can reduce
operational overhead but add to the overall cost. |
|
|
5.
Cost Optimization Strategies |
· To manage and
optimize costs when using Legal-BERT, consider the following strategies |
|
|
Use Spot or
Preemptible Instances |
· Both AWS and
Google Cloud offer these discounted instances, which are ideal for
non-time-sensitive workloads like model training. |
Scale Inference
on Demand |
· Use auto-scaling
to adjust the number of instances based on demand, reducing costs during
low-usage periods. |
Optimize Storage
Costs |
· Move less
frequently accessed data to cheaper storage tiers, such as Infrequent Access
(AWS), Nearline (Google Cloud), or Cool Storage (Azure). |
Reserved
Instances or Savings Plans |
· Commit to a
specific level of usage over time to benefit from significant cost savings,
especially for long-term projects. |
Construction
Contract Management Artificial Intelligence (AI) Services
AI
Knowledge Management Systems Limited
@Euro
Training Limited
Whatsapp +15512411304
LEGAL-BERT:
The Muppets straight out of Law School
Ilias Chalkidis, Manos Fergadiotis, Prodromos Malakasiotis, Nikolaos Aletras, Ion Androutsopoulos
BERT has achieved
impressive performance in several NLP tasks. However, there has been limited
investigation on its adaptation guidelines in specialised
domains. Here we focus on the legal domain, where we explore several approaches
for applying BERT models to downstream legal tasks, evaluating on multiple
datasets. Our findings indicate that the previous guidelines for pre-training
and fine-tuning, often blindly followed, do not always generalize well in the
legal domain. Thus we propose a systematic investigation
of the available strategies when applying BERT in specialised
domains. These are: (a) use the original BERT out of the box, (b) adapt BERT by
additional pre-training on domain-specific corpora, and (c) pre-train BERT from
scratch on domain-specific corpora. We also propose a broader hyper-parameter
search space when fine-tuning for downstream tasks and we release LEGAL-BERT, a
family of BERT models intended to assist legal NLP research, computational law,
and legal technology applications.
Comments: |
5 pages, short paper in Findings of EMNLP 2020 |
Subjects: |
Computation and Language (cs.CL) |
Cite as: |
arXiv:2010.02559 [cs.CL] |
|
(or arXiv:2010.02559v1 [cs.CL] for this version) |
|
https://doi.org/10.48550/arXiv.2010.02559 Focus to learn more |
Submission
history
From: Ilias
Chalkidis [view email]
[v1] Tue, 6 Oct 2020 09:06:07 UTC (249 KB)