cgato/Thespis-13b-v0.4

Model Overview

The cgato/Thespis-13b-v0.4 is a 13 billion parameter language model developed by cgato, designed primarily for multi-turn roleplay (RP) experiences. It is fine-tuned to produce both SFW and NSFW content, adhering to an internet-style RP format where actions are denoted by asterisks and dialogue is presented without quotes.

Key Capabilities

• Multi-turn Roleplay: Optimized for engaging in extended, dynamic character interactions.
• Internet-Style RP: Generates content consistent with common online roleplay conventions (e.g., *action* for actions, no quotes for dialogue).
• Content Flexibility: Capable of generating both SFW (safe for work) and NSFW (not safe for work) roleplay scenarios.

Training Data

The model was trained on a diverse dataset specifically curated for roleplay and conversational quality, including:

• 3,000 samples from the Claude Multiround Chat 30k dataset (filtered for length and excluding coding/math).
• 3,000 samples from the Pippa Dataset (filtered for length, asterisk usage, and lack of quotes).
• 2,200 samples from the Puffin Dataset (filtered for length and excluding coding/math).
• 700 samples from Airoboros 3.1 (long writing samples).
• 4,700 hand-curated RP conversations with various characters.

Usage

This model is compatible with standard chat formats for platforms like Oobabooga (Ooba) and SillyTavern. It uses a simple prompt format:

{System Prompt}

Username: {Input}
BotName: {Response}
Username: {Input}
BotName: {Response}

For Ooba Instruct, a specific turn template is provided:

<|user|>{Username}: <|user-message|>\n<|bot|>{BotName}: <|bot-message|>\n