Wayfarer-12B

Maintained By
LatitudeGames

Wayfarer-12B

PropertyValue
AuthorLatitudeGames
Base ModelNemo
Model URLhttps://huggingface.co/LatitudeGames/Wayfarer-12B
Training FormatChatML

What is Wayfarer-12B?

Wayfarer-12B is a specialized language model designed specifically for adventure role-playing experiences. Unlike conventional AI models that tend to be overly positive and protective of users, Wayfarer-12B introduces genuine challenge and consequences into narrative interactions. The model was developed by LatitudeGames to address the common criticism that modern AI models are too "nice" and lack the ability to create engaging conflict.

Implementation Details

The model employs a two-stage SFT (Supervised Fine-Tuning) approach, built on the Nemo architecture. The first stage involved 180K chat-formatted instruct data instances, while the second stage utilized a 50/50 mix of synthetic 8K context text adventures and roleplay experiences. The training data was generated through simulated playthroughs of AI Dungeon scenarios, featuring five distinct user archetypes to ensure variety.

  • Optimized for second-person present tense narration
  • Implements ChatML format for interactions
  • Recommended settings: temperature 0.8, repetition penalty 1.05, min_p 0.025
  • 8K context window for extended narrative sequences

Core Capabilities

  • Generates challenging and potentially fatal narrative scenarios
  • Creates genuine stakes and consequences in storytelling
  • Maintains pessimistic sentiment to counter typical AI positivity bias
  • Supports varied character creation and scenario development
  • Enables dynamic narrative responses without plot armor

Frequently Asked Questions

Q: What makes this model unique?

Wayfarer-12B stands out for its deliberate focus on challenging gameplay and narrative consequences. Unlike most AI models that are trained to be helpful and protective, this model embraces failure states and genuine risk, making it ideal for authentic gaming experiences.

Q: What are the recommended use cases?

The model is best suited for text-based adventure games, role-playing scenarios, and interactive fiction where genuine challenge and consequences are desired. It excels in second-person present tense narratives and is optimized for single-turn chat interactions.

🍰 Interesting in building your own agents?
PromptLayer provides Huggingface integration tools to manage and monitor prompts with your whole team. Get started here.