Dans-Archive/Dans-CreepingSenseOfDoom-13b

TEXT GENERATIONConcurrency Cost:1Model Size:13BQuant:FP8Ctx Length:4kPublished:Aug 25, 2023Architecture:Transformer0.0K Cold

Dans-Archive's Dans-CreepingSenseOfDoom-13b is a 13 billion parameter language model, fine-tuned from KoboldAI's LLAMA2-13B-Holodeck-1, specifically designed for generating text-based adventure games. It excels at crafting interactive narratives, capable of producing both concise and expansive, novel-like descriptions. The model leverages a 4096-token context length and is optimized for dynamic story generation based on user input and system messages.

Loading preview...

Model Overview

Dans-CreepingSenseOfDoom-13b is a 13 billion parameter model developed by Dans-Archive, fine-tuned from KoboldAI's LLAMA2-13B-Holodeck-1. Its primary function is to act as a game master for text-based adventure games, generating interactive narratives based on user actions and predefined themes. The model was trained using a deduped version of the Skein text adventure dataset, augmented with system messages following the 'Metharme' prompting format.

Key Capabilities

  • Text-Based Adventure Game Generation: Proficient in creating dynamic and interactive story experiences.
  • Variable Response Lengths: Can produce both concise replies and detailed, novel-like descriptions, controllable via system messages.
  • Flexible Thematic Adaptation: Guides stories based on diverse themes (e.g., Science Fiction, Horror, Fantasy).
  • Second-Person Present Tense Narration: Effectively narrates actions and observations from the player's perspective.

Training Details

The model was trained using QLoRA on a single RTX 3090 for 3 epochs, taking approximately 8 hours. It utilizes a sequence length of 4096 tokens. The training process incorporated the 'Pygmalion Metharme' prompt format, allowing for distinct system messages to influence response style and content.

Good For

  • Developers building interactive fiction platforms or text adventure games.
  • Creating dynamic narrative experiences where the model acts as a dungeon master or storyteller.
  • Generating varied response lengths in conversational AI, from short actions to descriptive passages.