Styled Lines

This is a GPT2-based line rewritter.

We ship it with three model quantised resolutions, try it online entirely in your browser with the quantised (simplified) one (Q4_K_M 110mb):

- Automated NPC Dialogue stylisation demo

- Single Prompt Debug demo

Inside the package, you will find Q8_0 (175mb) and original bf16 (321mb) resolution models ready to run inside a cross-platform game engine!

You give it:

```

<input> How are you today? <inputEnds>

<style> Pirate's Poetic Question <styleEnds>

<output>

```

It prints out: Hey there, how fares ye today?

We hope that this model and wrapper can help you save a day! (or, more realistically, 3 months with a dedicated team of 3 people in our estimates).

• Model finetuned on a corpus of more than 0.5 million dialogue lines (64343812 tokens).
• All our training data was purely syntactical and generated.
• Tools for output filtering are provided with release in CSharp - thus easely editable!
• All generative LLM models can make mistakes in their output. In some cases, they can turn questions into statements and add minor hallucinated info based on the input text and style data.

The bounded model inference wrapper is based on open-source model runtime and supports the latest 405b models, so you can try them out on platforms with enough RAM/time available.

The bundled wrapper allows:

• Build and tested for: Web, Mobile, PC.
• To execute model setup and execution in asynchronous mode based on threads in C++ code. (this is needed if one wants parallel code run when building targeting the Web platform)
• It allows tapping into Logs, output Tokens, and completion callbacks from generation.
• Setup model update callbacks in editor UI using events and helper scripts. Also, you can configure most model configuration/run parameters from the editor UI and at runtime.