Model Overview

The cgato/Thespis-CurtainCall-7b-v0.3 is a 7 billion parameter language model developed by cgato, featuring an 8192 token context window. This iteration is a fine-tuned model, building upon a diverse set of training data to enhance its conversational and reasoning capabilities.

Key Training Datasets

The model was trained on a comprehensive collection of datasets, indicating a focus on broad conversational understanding and specific reasoning tasks. Notable datasets include:

• Dolphin, Ultrachat, Capybara, Airoboros 3.1, OpenOrca: These contribute to general conversational fluency and instruction following.
• Augmental, Yahoo Answers: Likely enhance its ability to process and generate informative responses.
• ToxicQA: Suggests an effort to handle or understand potentially sensitive or challenging content.
• Magiccoder-Evol-Instruct-110k: Indicates some exposure to code-related instructions or logical reasoning.
• grimulkan/physical-reasoning and theory-of-mind: These datasets are particularly interesting, pointing towards an emphasis on understanding physical interactions and inferring mental states, which can be crucial for advanced conversational agents and role-playing scenarios.

Recommended Usage

The model is optimized for chat-based interactions and follows a standard chat prompt format. It is compatible with popular interfaces like Oobabooga and Silly Tavern. Recommended settings for optimal performance in Silly Tavern include a Temperature of 1.25, MinP of 0.1, and Repetition Penalty of 1.03.