dreamgen/opus-v1-34b

Overview

dreamgen/opus-v1-34b is a 34 billion parameter model from the DreamGen Opus V1 series, specifically designed and fine-tuned for steerable story-writing and role-playing. It employs an extended ChatML prompting format, introducing a text role with optional names to indicate active characters, allowing for nuanced control over narrative flow and character dialogue. The model was trained on approximately 100 million tokens of diverse examples, including human-written story-writing, role-playing, writing assistance, and general assistance data, with individual examples up to 31,000 tokens long.

Key Capabilities

• Steerable Story-writing and Role-playing: Generates continuations based on detailed system prompts (plot, style, characters) and user instructions.
• Story Plot Summarization: Condenses stories or chapters into concise plot descriptions of varying lengths.
• Story Character Description: Extracts and describes characters from provided story texts.
• Story Style Description: Analyzes and describes the writing style of a given story.
• Story Description to Chapters: Breaks down a brief plot description into chapter-by-chapter outlines.

Good For

• Interactive Fiction: Creating dynamic and responsive story experiences.
• Creative Writing Assistance: Aiding authors with plot development, character creation, and style consistency.
• Role-playing Scenarios: Generating engaging and steerable character interactions.
• Content Generation: Producing summaries, descriptions, and outlines for narrative content.

DreamGen recommends specific sampling parameters for optimal performance in story-writing and role-play, suggesting min_p between 0.01 and 0.1, and temperature between 0.5 and 1.5, with min_p=0.1; temperature=0.8 as a good starting point. Users should pay close attention to the extended ChatML prompting format and tokenization for best results.