Getting Started with Marqtune
This page contains information about how to get up and running with Marqtune.
These are the high level steps to fine-tune your embedding model. Details are described in the associated guides and API reference.
- Create CSV files and Marqtune datasets for training and evaluation
- Select a base model and execute a model training task with your training dataset
- View the training logs via the UI or API as the task is executing and when complete
- Evaluate the model's checkpoints using your evaluation dataset to determine the best performing epoch
- Host the checkpoint file at a publicly accessible URL
- Release the model checkpoints for inference on a Marqo Cloud index
Sign up to Marqo Cloud and Request Marqtune Access
Sign up at the Marqo Cloud console, select Marqtune from the left-hand navigation and request access. If your request for access is approved, a Marqo team member will contact you via your Marqo Cloud registered email address and grant access on your account.
Create Training and Evaluation Datasets
A training dataset is a required input for a Marqtune training task. We recommend an 80/20 split; 80% of your data for the training dataset and 20% for an evaluation dataset.
You can read more about the data requirements and about using the API or the Marqtune Python client creating a dataset for training and evaluation.
Select a Model to Fine-Tune
You can choose from amongst a large number of models to fine-tune on Marqtune. Notably amongst them, we recommend the following OpenCLIP image+text models: - ViT-B-32 - ViT-L-14
And the following text-only models: - intfloat/e5-v2-small - intfloat/e5-v2-large
Choose a Marqtune Instance Type
| Name | Description | Per minute price (USD) |
|---|---|---|
| marqtune.basic | Low cost, recommended for dev/testing workloads | $0.0202 |
| marqtune.performance | Performance optimised for production workloads | $0.6282 |
marqtune.basic Marqtune basic is the cheapest Marqtune instance available. It is targeted towards testing and development.
marqo.performance Marqtune performance is the best performing Marqtune instance type. This is suitable for production applications.
As a general rule, the larger the batch size used for fine-tuning, the better the outcome. The choice of Marqtune instance type depends on the size of the base model you are training and the batch size required. We recommend testing Marqtune using the basic instance using smaller batch sizes, and then moving to a performance instance and larger batch sizes for your production ready model.
Getting Started with Marqtune
Follow the Fine-Tuning Multimodal Embedding Models: End to End Walkthrough which will guide you through the process of creating datasets, training, and evaluation before finally releasing the model to Marqo Cloud or downloading a tuned model for use as a custom model for your Marqo index.
How do I Use a Fine-Tuned Model in Marqo?
Marqo supports using custom weights for the embedding model. You can release a model's checkpoints to Marqo Cloud. Alternatively you can load the fine tuned model yourself at index creation time in Marqo.
What is the Cost?
Marqtune is billed per minute for training and evaluation tasks only. See the Marqtune pricing guide for details.
Who Owns the Fine-Tuned Model?
You own the model and its fine-tuned weights. The model's checkpoints can be downloaded via the Marqtune API.
How Long Does it Take?
The time it takes depends on the model size, dataset size, data types, unique images, batch size, number of workers and the Marqtune instance type used. The training task can take many hours depending on the dataset size.
As an example, it took just over 4 hours (4hrs 3min 32sec) to fine-tuned ViT-B-32:laion400m_e31 for 20 epochs, with a 256 batch size, 2 workers and a 250k record dataset. The dataset was an ideal Marqtune dataset; it contained the columns query, product image, product title and score. It had approximatly 100k unqiue images and was trained using instance type marqtune.basic.
Will our Training Data be Used for Other Purposes?
No. Your data will not be used by Marqo for any purpose other than supplying the Marqo Cloud and Marqtune service to you, in accordance with our Marqo Cloud Terms and Conditions.