Wayfarer-2-12B: Challenging AI Storytelling
Wayfarer-2-12B is a 12 billion parameter model developed by LatitudeGames, designed to create dynamic and challenging text-adventure experiences. It addresses the common issue of AI models being "too nice" by intentionally incorporating conflict, tension, and the possibility of character failure or death, enhancing narrative stakes.
Key Capabilities & Features
- Pessimistic Narrative Generation: Specifically trained to produce stories with frequent failure and a lack of "plot armor" for characters, including the user.
- Enhanced Detail & Pacing: Delivers longer, more detailed responses and a slower narrative pace compared to its predecessor.
- Text Adventure Optimization: Fine-tuned using simulated playthroughs of AI Dungeon scenarios, covering diverse user archetypes and character starts.
- Sentiment-Balanced Training: Utilizes a unique dataset combining text adventure data, sentiment-balanced roleplay transcripts, and a small instruct core.
- Second-Person Present Tense: Primarily trained on and optimized for narratives in second-person present tense (e.g., "You are").
How It Was Made
The model underwent SFT training using a three-ingredient recipe: the Wayfarer 2 dataset, sentiment-balanced roleplay transcripts, and an instruct core. The text adventure data was generated by having two language models (one narrator, one user) simulate playthroughs of AI Dungeon scenarios, running for up to 8k tokens or until character death. This process ensured a high degree of unpredictability and challenge.
Recommended Use Cases
- Interactive Fiction & Text Adventures: Ideal for developers creating games or applications that require a challenging, high-stakes narrative experience.
- Role-Playing Games (RPGs): Suitable for generating dynamic GM responses where player actions have significant, often perilous, consequences.
- Narrative Generation: For projects seeking to break from overly positive or predictable AI-generated stories, offering a more gritty and realistic tone.
Limitations
While other perspectives may work, the model is primarily optimized for second-person present tense narratives, and using different perspectives might yield suboptimal results.