Training Data Format

Overview#

Rasa uses YAML as a unified and extendable way to manage all training data, including NLU data, stories and rules. Rasa Studio provides an additional layer on top of that, enabling the management of training data through a web-based interface.

You can split the training data over any number of YAML files, and each file can contain any combination of NLU data, stories, and rules. The training data parser determines the training data type using top level keys.

The domain uses the same YAML format as the training data and can also be split across multiple files or combined in one file. The domain includes the definitions for responses and forms. See the documentation for the domain for information on how to format your domain file.

High-Level Structure#

Each file can contain one or more keys with corresponding training data. One file can contain multiple keys, but each key can only appear once in a single file. The available keys are:

• version
• nlu
• stories
• rules

You should specify the version key in all YAML training data files. If you don't specify a version key in your training data file, Rasa will assume you are using the latest training data format specification supported by the version of Rasa you have installed. Training data files with a Rasa version greater than the version you have installed on your machine will be skipped. Currently, the latest training data format specification for Rasa 3.x is 3.1.

Example#

Here's a short example which keeps all training data in a single file:

To specify your test stories, you need to put them into a separate file:

Test stories use the same format as the story training data and should be placed in a separate file with the prefix test_.

NLU Training Data#

NLU training data consists of example user utterances categorized by intent. Training examples can also include entities. Entities are structured pieces of information that can be extracted from a user's message. You can also add extra information such as regular expressions and lookup tables to your training data to help the model identify intents and entities correctly.

NLU training data is defined under the nlu key. Items that can be added under this key are:

• Training examples grouped by user intent e.g. optionally with annotated entities

• Synonyms

• Regular expressions

• Lookup tables

Training Examples#

Training examples are grouped by intent and listed under the examples key. Usually, you'll list one example per line as follows:

However, it's also possible to use an extended format if you have a custom NLU component and need metadata for your examples:

The metadata key can contain arbitrary key-value data that is tied to an example and accessible by the components in the NLU pipeline. In the example above, the sentiment metadata could be used by a custom component in the pipeline for sentiment analysis.

You can also specify this metadata at the intent level:

In this case, the content of the metadata key is passed to every intent example.

If you want to specify retrieval intents, then your NLU examples will look as follows:

All retrieval intents have a suffix added to them which identifies a particular response key for your assistant. In the above example, ask_name and ask_weather are the suffixes. The suffix is separated from the retrieval intent name by a / delimiter.

Entities#

Entities are structured pieces of information that can be extracted from a user's message.

Entities are annotated in training examples with the entity's name. In addition to the entity name, you can annotate an entity with synonyms, roles, or groups.

In training examples, entity annotation would look like this:

The full possible syntax for annotating an entity is:

The keywords role, group, and value are optional in this notation. The value field refers to synonyms. To understand what the labels role and group are for, see the section on entity roles and groups.

Synonyms#

Synonyms normalize your training data by mapping an extracted entity to a value other than the literal text extracted. You can define synonyms using the format:

You can also define synonyms in-line in your training examples by specifying the value of the entity:

Read more about synonyms on the NLU Training Data page.

Regular Expressions#

You can use regular expressions to improve intent classification and entity extraction using the RegexFeaturizer and RegexEntityExtractor components.

The format for defining a regular expression is as follows:

Here account_number is the name of the regular expression. When used as features for the RegexFeaturizer the name of the regular expression does not matter. When using the RegexEntityExtractor, the name of the regular expression should match the name of the entity you want to extract.

Read more about when and how to use regular expressions with each component on the NLU Training Data page.

Lookup Tables#

Lookup tables are lists of words used to generate case-insensitive regular expression patterns. The format is as follows:

When you supply a lookup table in your training data, the contents of that table are combined into one large regular expression. This regex is used to check each training example to see if it contains matches for entries in the lookup table.

Lookup table regexes are processed identically to the regular expressions directly specified in the training data and can be used either with the RegexFeaturizer or with the RegexEntityExtractor. The name of the lookup table is subject to the same constraints as the name of a regex feature.

Read more about using lookup tables on the NLU Training Data page.

Conversation Training Data#

Stories and rules are both representations of conversations between a user and a conversational assistant. They are used to train the dialogue management model. Stories are used to train a machine learning model to identify patterns in conversations and generalize to unseen conversation paths. Rules describe small pieces of conversations that should always follow the same path and are used to train the RulePolicy.

Stories#

Stories are composed of:

• story: The story's name. The name is arbitrary and not used in training; you can use it as a human-readable reference for the story.
• metadata: arbitrary and optional, not used in training, you can use it to store relevant information about the story like e.g. the author
• a list of steps: The user messages and actions that make up the story

For example:

Each step can be one of the following:

• A user message, represented by intent and entities.
• An or statement, which includes two or more user messages under it.
• A bot action.
• A form.
• A slot was set event.
• A checkpoint, which connects the story to another story.

User Messages#

All user messages are specified with the intent: key and an optional entities: key.

While writing stories, you do not have to deal with the specific contents of the messages that the users send. Instead, you can take advantage of the output from the NLU pipeline, which uses a combination of an intent and entities to refer to all possible messages the users can send with the same meaning.

User messages follow the format:

For example, to represent the sentence I want to check my credit balance, where credit is an entity:

It is important to include the entities here as well because the policies learn to predict the next action based on a combination of both the intent and entities (you can, however, change this behavior using the use_entities attribute).

Actions#

All actions executed by the bot are specified with the action: key followed by the name of the action. While writing stories, you will encounter two types of actions:

1. Responses: start with utter_ and send a specific message to the user. e.g.

2. Custom actions: start with action_, run arbitrary code and send any number of messages (or none).

Forms#

A form is a specific kind of custom action that contains the logic to loop over a set of required slots and ask the user for this information. You define a form in the forms section in your domain. Once defined, you should specify the happy path for a form as a rule. You should include interruptions of forms or other conversation repair in stories so that the model can generalize to unseen conversation sequences. As a step in a story, a form takes the following format:

The action step activates the form and begins looping over the required slots. The active_loop: restaurant_form step indicates that there is a currently active form. Much like a slot_was_set step, a form step doesn't set a form to active but indicates that it should already be activated. In the same way, the active_loop: null step indicates that no form should be active before the subsequent steps are taken.

A form can be interrupted and remain active; in this case the interruption should come after the action: <form to activate> step and be followed by the active_loop: <active form> step. An interruption of a form could look like this:

Slots#

A slot event is specified under the key slot_was_set: with the slot name and optionally the slot's value.

Slots act as the bots memory. Slots are set by either the default action action_extract_slots according to the slot mappings specified in the domain, or by custom actions. They are referenced by stories in slot_was_set steps. For example:

This means the story requires that the current value for the feedback_value slot be positive for the conversation to continue as specified.

Whether or not you need to include the slot's value depends on the slot type and whether the value can or should influence the dialogue. If the value doesn't matter, as is the case for e.g. text slots, you can list only the slot's name:

The initial value for any slot by default is null, and you can use it to check if the slot was not set:

Checkpoints#

Checkpoints are specified with the checkpoint: key, either at the beginning or the end of a story.

Checkpoints are ways to connect stories together. They can be either the first or the last step in a story. If they are the last step in a story, that story will be connected to each other story that starts with the checkpoint of the same name when the model is trained. Here is an example of a story that ends with a checkpoint, and one that starts with the same checkpoint:

Checkpoints at the beginning of stories can also be conditional on slots being set, for example:

Checkpoints can help simplify your training data and reduce redundancy in it, but do not overuse them. Using lots of checkpoints can quickly make your stories hard to understand. It makes sense to use them if a sequence of steps is repeated often in different stories, but stories without checkpoints are easier to read and write.

OR statement#

or steps are ways to handle multiple intents or slot events the same way, without writing a separate story for each intent. For example, if you ask the user to confirm something, you might want to treat the affirm and thankyou intents in the same way. Stories with or steps will be converted into multiple separate stories at training time. For example, the following story would be converted to two stories at training time:

You can also use or statements with slot events. The following means the story requires that the current value for the name slot is set and is either joe or bob. This story would be converted to two stories at training time.

Just like checkpoints, OR statements can be useful, but if you are using a lot of them, it is probably better to restructure your domain and/or intents.

Rules#

Rules are listed under the rules key and look similar to stories. A rule also has a steps key, which contains a list of the same steps as stories do. Rules can additionally contain the conversation_started and conditions keys. These are used to specify conditions under which the rule should apply.

A rule that with a condition looks like this:

For more information about writing rules, see Rules.

Test Stories#

Test stories check if a message is classified correctly as well as the action predictions.

Test stories use the same format as stories, except that user message steps can include a user to specify the actual text and entity annotations of the user message. Here's an example of a test story:

You can run the tests using the following command:

If you want to know more about testing head over to Testing Your Assistant.

End-to-end Training#

With end-to-end training, you do not have to deal with the specific intents of the messages that are extracted by the NLU pipeline. Instead, you can put the text of the user message directly in the stories, by using user key.

These end-to-end user messages follow the format:

In addition, you can add entity tags that can be extracted by the TED Policy. The syntax for entity tags is the same as in the NLU training data. For example, the following story contains the user utterance I can always go for sushi. By using the syntax from the NLU training data [sushi](cuisine), you can mark sushi as an entity of type cuisine.

Similarly, you can put bot utterances directly in the stories, by using the bot key followed by the text that you want your bot to say.

A story with only a bot utterance might look like this:

You can also have a mixed end-to-end story:

Rasa end-to-end training is fully integrated with standard Rasa approach. It means that you can have mixed stories with some steps defined by actions or intents and other steps defined directly by user messages or bot responses.