Creating Entries // RoleThread Lite Docs

Create Entry is where you write new training examples for the loaded dataset. It is designed for a steady writing flow: set the system prompt, write one or more user/assistant exchanges, add tags, preview the result, then save.

You do not need to make every entry perfect on the first pass. RoleThread Lite gives you editing, validation, search, duplicate, split, and join tools so you can refine entries over time.

Before You Start

Create Entry works after a dataset is loaded.

If no dataset is loaded, go to Manage Dataset first and either:

create a new dataset
load an existing JSONL file
load a working copy created by RoleThread

New entries are saved into the currently loaded dataset.

The Basic Entry Shape

Most entries contain:

A system message that sets the context.
One or more user messages.
One or more assistant messages.

The user and assistant messages are written as exchange pairs. A single-turn entry has one user message and one assistant response. A multi-turn entry has multiple pairs under the same system prompt.

Example shape:

{
  "messages": [
    {"role": "system", "content": "You are writing a grounded character scene."},
    {"role": "user", "content": "Ask about the locked door."},
    {"role": "assistant", "content": "Assistant studies the frame, then lowers her voice..."}
  ]
}

RoleThread handles the JSON structure for you. You write the content.

System Prompt

The system prompt tells the training example what kind of response belongs in the entry.

Good system prompts are usually:

specific enough to set the scene or style
short enough to stay focused
consistent with the entry content
reusable when several entries share the same setup

You can write a custom prompt directly, or use Load from library to fill the field from a saved system prompt template. Loading from the library does not lock the field. You can still edit the prompt before saving.

If you write a prompt you expect to reuse, use Save as template to add it to the System Prompts library.

Exchanges

Each exchange has a user side and an assistant side.

Use the planned exchange count to decide how many pairs you want in the entry. The save button stays disabled until the planned exchanges are complete and valid.

Use multiple exchanges when:

the scene needs a short back-and-forth
the assistant response depends on earlier user turns
you want to train continuity within a focused moment

Avoid making one entry carry too many unrelated beats. If an entry grows into multiple scenes, the Split tool in Full Edit may be a better fit later.

Validation While Writing

Create Entry shows a preview of the entry before saving. If the entry is valid, RoleThread tells you it looks valid.

Entries created through the normal Create Entry flow are usually structurally safe by default because RoleThread validates the shape before save. Validation is still useful later for review, cleanup, and consistency checks across the whole dataset.

The save button may be disabled when:

planned exchanges are incomplete
required message content is blank
the entry has validation problems
no dataset is loaded

When a button is disabled, read the small caption near it. It usually tells you what still needs to be finished.

Entry Quality Mindset

Good training entries are not only valid JSON. They should also be useful examples.

Aim for:

clear setup
focused scene intent
assistant responses with enough detail to teach style and behavior
consistent tone within the entry
no accidental refusal or meta-language unless that is truly part of the dataset
no unrelated conversations stitched together

For many narrative datasets, assistant responses that are too short are less useful than responses that show the behavior you want the model to learn.

Writing Effective Narrative Training Data

Writing choices teach patterns.

That does not mean there is one correct style. It means repeated structure, wording, formatting, and point of view can influence how a trained model behaves.

Narrative-Heavy vs Dialogue-Heavy

A narrative-heavy dataset may encourage more descriptive, atmospheric responses.

A dialogue-heavy dataset may encourage more conversational, back-and-forth responses.

Neither is automatically better. Choose the balance that matches the model behavior you want.

First Person and Third Person

Keep point of view consistent within a dataset section when you can.

If some entries use first person and others use third person, that can be fine, but the model may learn to mix styles unless the difference is intentional.

Names and Pronouns

Repeated character names can teach strong character-specific associations.

That is useful in some Group Chat scenes, but it can make a behavior feel tied to one named character instead of generalizing across similar situations.

Example:

Assistant walks over to the table and smiles at User.

This is clear, but repeated name-heavy narration can over-bind the behavior to Assistant and User.

For more reusable behavior, pronouns may generalize better:

She walks over to the table and smiles at him.

Use names when they clarify who is acting, especially in group scenes. Use pronouns when the behavior matters more than the specific names.

Quoted Dialogue

Consistent quoted dialogue helps teach conversational structure.

Example:

"I don't think this is a good idea," she says quietly.

If your dataset mixes quoted dialogue, unquoted dialogue, markdown bullets, and script format without intention, the model may learn that inconsistency.

Formatting Consistency

Formatting is part of the example.

If assistant responses repeatedly use a certain style, the model may reproduce it. This includes:

quotation style
paragraph breaks
markdown
stage directions
name labels
tense and point of view

Use formatting intentionally. Consistency teaches consistency.

Practical Workflow

A simple creation workflow:

Choose Default or Group Chat mode.
Write or load a system prompt.
Set the planned exchange count.
Fill each user/assistant exchange.
Add relevant tags.
Review the JSON preview.
Save the entry.
Use Manage Dataset to review, filter, tag, duplicate, or quick edit entries as the dataset grows.
Run Validation when you want a broader review pass, especially before export.

Common Mistake

Mistake: Trying to make one entry cover an entire long scene.

Better approach: Keep entries focused. If a conversation naturally has several distinct moments, consider multiple entries or use Full Edit to split it later.

What Happens After Save

When an entry is saved, RoleThread assigns stable identity metadata such as an entry UUID, updates the dataset, refreshes sidecar metadata, and keeps the entry available for search, editing, validation, and export.

You can always return to the entry later through Manage Dataset or Deep Edit.