maldv/dragonwar-7b-alpha

Hugging Face
TEXT GENERATIONConcurrency Cost:1Model Size:7BQuant:FP8Ctx Length:8kPublished:Apr 5, 2024License:cc-by-nc-4.0Architecture:Transformer0.0K Open Weights Warm

maldv/dragonwar-7b-alpha is a 7 billion parameter language model specifically trained for generating fantasy story content. It utilizes a unique training regimen involving full text windows, ORPO, and rotated epochs to enhance coherence in narrative generation. This model is optimized for creative writing in storymode applications, offering a rich cast of characters from various fantasy universes.

Loading preview...

Dragonwar 7b - α Overview

maldv/dragonwar-7b-alpha is a 7 billion parameter language model designed for generating immersive fantasy narratives. Its training methodology involved full text windows, followed by completion, then ORPO (Odds Ratio Preference Optimization), and a final epoch of full text with a 1/4 window rotation. This unique approach aims to produce highly coherent and engaging story content.

Key Capabilities

  • Fantasy Story Generation: Optimized for creating detailed and coherent fantasy narratives.
  • Rich Character Integration: Capable of drawing from a wide array of established fantasy characters (e.g., King Robb Stark, Lord Rahl, Perrin, Zeddicus, Gandalf).
  • Author's Notes Support: Effectively utilizes author's notes to guide story direction and character interactions.
  • No EOS Tokens: A specific hack was implemented to remove End-Of-Sentence (EOS) tokens, ensuring continuous generation without premature stops.

Intended Use Cases

This model is not a chat model but is specifically intended for:

  • Storymode Applications: Generating ongoing story segments or entire narratives.
  • Creative Writing: Assisting writers in developing fantasy plots and character interactions.
  • Interactive Fiction: Powering text-based adventure games or interactive storytelling platforms.

Users are advised to start with a bit of story or a character name rather than a traditional prompt. Recommended generation settings include Min-P of 0.1, dynatemp between 0.5 and 2, and smoothing between 0.05 and 0.2.

Popular Sampler Settings

Top 3 parameter combinations used by Featherless users for this model. Click a tab to see each config.

temperature
top_p
top_k
frequency_penalty
presence_penalty
repetition_penalty
min_p