SlerpE/CardProjector-14B-v2

TEXT GENERATIONConcurrent Unit Cost:1Model Size:14.8BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Mar 10, 2025License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

CardProjector-14B-v2 by WoonaAi is a 14.8 billion parameter language model fine-tuned for generating detailed character cards. This model excels at crafting rich, consistent, and engaging character descriptions for roleplaying, specifically optimized for SillyTavern and general character creation. It features a redesigned output format that abandons JSON for improved quality, multi-turn conversation support, and creative writing flexibility, allowing for high temperatures without output breakage.

Loading preview...

Overview

CardProjector-14B-v2, developed by WoonaAi, is a 14.8 billion parameter language model specifically fine-tuned for generating detailed character cards. This second version represents a significant update, moving away from a JSON output format to enhance output quality, improve multi-turn conversation for character editing, and provide greater flexibility for creative writing.

Key Capabilities

  • Enhanced Character Generation: Produces significantly improved, more detailed, consistent, and engaging character descriptions, backstories, and personalities.
  • Flexible Output: The redesigned output format allows for creative freedom, supporting high temperatures (up to 1.1) without breaking the output structure.
  • Multi-Turn Editing: Improved ability to handle edit requests, leading to more consistent character map modifications.
  • SillyTavern & General Character Creation: Designed to assist creators and roleplayers in crafting characters compatible with SillyTavern, and now also suitable for general character creation.
  • Improved Logical Component: Features an enhanced logical understanding compared to previous versions.

Usage Recommendations

  • Chat Template: Utilizes ChatML for optimal performance.
  • Configurable Output: Provides specific temperature, Top-P, Rp.Pen, Top-K, Rp.Slope, and Rp.Range settings for balanced output, bright descriptions, and best instruction following.

Limitations

  • Currently, the model does not support generating character cards in SillyTavern JSON format, only natural text. Future versions aim to convert output to JSON based on dialogue history.
  • The 14B model may exhibit some issues with instruction following and can be overly reliant on specific response formats, which are targeted for future fixes.

Finetunes
2 models