Thespis-Mistral-7b-v0.6 Overview

The cgato/Thespis-Mistral-7b-v0.6 is a 7 billion parameter language model built upon the Mistral architecture. This version, developed by cgato, is a personal project specifically fine-tuned for multi-turn roleplay (RP), supporting both SFW and NSFW scenarios. It is noted to perform comparably to models between Llama 7B and Llama 13B in internal testing.

Key Capabilities & Features

• Roleplay Optimization: Designed for immersive, multi-turn roleplay experiences.
• Internet-Style RP: Best suited for RP using standard markup (asterisks for actions, no quotes for dialogue).
• Diverse Training Data: Fine-tuned on a unique combination of datasets, including:
  • 2200 samples from Pure-Dove Dataset (>= 90 token length)
  • 2200 samples from Claude Multiround 30k (>= 90 token length)
  • 700 samples from Airoboros 3.1 (writing samples > 1500 tokens)
  • 900 samples from the Augmental Dataset (>= 90 token length)
  • 6000 samples of hand-curated RP conversations.
• Standard Chat Format: Compatible with default chat templates for platforms like Oobabooga and SillyTavern.

Good for

• Developers and users seeking a specialized model for interactive, character-driven roleplay.
• Applications requiring a model that understands and generates internet-style RP dialogue and actions.
• Integration with frontends like Oobabooga and SillyTavern for a seamless RP experience.