# Training pipelines

> :::important
Minimal dataset size For successfully running a Training pipeline, we strongly recommend minimum 10 documents and at least 5 samples from each labeled field in your dataset. Otherwise, the pipeline throws the following error: `Dataset Creation Failed`.

:::important
Minimal dataset size For successfully running a Training pipeline, we strongly recommend minimum 10 documents and at least 5 samples from each labeled field in your dataset. Otherwise, the pipeline throws the following error: `Dataset Creation Failed`.

Training on GPU vs CPU For larger datasets, you need to train using GPU. Moreover, using a GPU for training is at least 10 times faster than using a CPU. For the maximum dataset size depeding on the version and infrastructure, check the following table. Table 1. Maximum dataset for each version
| Infrastructure | <2021.10.x | 2021.10.x | >2021.10.x |
| --- | --- | --- | --- |
| **CPU** | 500 pages | 5000 pages | 1000 pages |
| **GPU** | 18,000 pages | 18,000 pages | 18,000 pages |

For more information on dataset structure, check the [Dataset format](https://docs.uipath.com/document-understanding/automation-suite/2024.10/classic-user-guide/document-manager-export-documents#dataset-format) section.
:::

There are two ways to train an ML model:

* training a model from scratch
* retraining an out-of-the-box model

Training a model from scratch can be done using the **DocumentUnderstanding** ML package, which does this on the dataset provided as input.

Retraining can be done using out-of-the-box ML packages such as **Invoices**, **Receipts**, **Purchase Orders**, **Utility Bills**, **Invoices India**, **Invoices Australia** etc., basically, any other data extraction ML Package except for **DocumentUnderstanding**. Training using one of these packages has one additional input: a base model. We refer to this as **retraining** because you are not starting from scratch but from a base model. This approach uses a technique called **Transfer Learning** where the model takes advantage of the information encoded in another model, the preexisting one. The model retains some of the out-of-the-box knowledge, but it learns from the new data too. However, as your training dataset size increases, the pretrained base model matters less and less. It is mainly relevant for small to medium size training datasets (up to 500-800 pages).

Configure the training pipeline as follows:

* In the **Pipeline type** field, select **Train run**.
* In the **Choose package** field, select the package you created based on the **DocumentUnderstanding** ML Package.
* In the **Choose package major version** field, select a major version for your package.
* In the **Choose package minor version** field, select a minor version for your package. It is strongly recommended to always use minor version 0 (zero). Check the **Choosing the minor version** section for more information.
* In the **Choose input dataset** field, select a dataset. For building high quality training datasets, you can check [this](https://docs.uipath.com/document-understanding/automation-suite/2024.10/classic-user-guide/training-high-performing-models#training-high-performing-models) tutorial.
* In the **Enter parameters** section, enter any environment variables defined, and used by your pipeline, if any. For most use cases, no parameter needs to be specified; the model is using advanced techniques to find a performant configuration. However, here are some environment variables you could use:
* `model.epochs` which customizes the number of epochs for the Training Pipeline (the default value is 100).
  :::note
  For larger datasets, containing more than 5000 pages, you can initially perform a full pipeline run with the default number of epochs. This allows you to evaluate the model’s accuracy. After that, you can decrease the number of epochs to about 30-40. This approach allows you to compare the accuracy of the results and determine if the reduction of epochs yields comparable precision. When using smaller datasets, in particular those with fewer than 5000 pages, you can maintain the default number of epochs.
  :::
* Select whether to train the pipeline on GPU or on CPU. The **Enable GPU** slider is disabled by default, in which case the pipeline is trained on CPU.
* Select one of the options when the pipeline should run: **Run now**, **Time based** or **Recurring**. In case you are using the `auto_retraining` variable, select **Recurring**.

  ![Screenshot of the Create new pipeline run interface.](https://dev-assets.cms.uipath.com/assets/images/document-understanding/document-understanding-screenshot-of-the-create-new-pipeline-run-interface-115811-8daa5512-bce1a718.webp)
* After you configure all the fields, select **Create**. The pipeline is created.

Here is an example of creating a new Training Pipeline with a dataset previously exported to AI Center:

## Choosing the minor version

In most situations, minor version **0** should be chosen. This is because the larger and more diverse your training dataset, the better your model's performance. This principle aligns with the current state-of-the-art ML technology's goal of using large, high-quality, and representative training sets. Therefore, as you accumulate more training data for a model, you should add the data to the same dataset to further enhance the model's performance.

There are situations, however, where training on a minor version other than 0 makes sense. This is typically the case when a partner needs to service multiple customers in the same industry, but UiPath® doesn't have a pre-trained model optimized for that industry, geography, or document type.

In such a case, the partner might develop a pre-trained model using a variety of document samples from that industry (not from a single source, but from many for better generalization). This model would be used as a base model to train specific customer models, being trained on version **0** of the ML package. Following versions, like version **1**, would be used to refine either the pre-trained model or create customer-specific models.

However, to obtain good results, the pre-trained model should be unbiased and based on a highly diverse training set. If the base model is optimized for a specific customer, it may not perform well for other customers. In such a case, using the zero minor version as a base model yields better results.