Hugging Face Overview
Wayfarer-12B: An Unforgiving Adventure Role-Play Model
Wayfarer-12B, developed by LatitudeGames, is a 12 billion parameter model built upon the Nemo base architecture. It stands out by being specifically trained to deliver challenging and dangerous adventure role-play experiences, a direct response to the common feedback that many modern AI models are overly 'nice' and lack conflict or consequences.
Key Capabilities & Training
- Designed for Conflict: Unlike many models aligned away from 'darkness' or 'violence', Wayfarer embraces these elements to create engaging narratives with real stakes, including player failure and death.
- Two-Stage SFT: The model was fine-tuned using a two-stage Supervised Fine-Tuning (SFT) approach. The first stage involved 180K chat-formatted instruct data instances, followed by a second stage using a 50/50 mixture of synthetic 8k context text adventures and roleplay experiences.
- Pessimistic Sentiment: Wayfarer's general emotional sentiment is one of pessimism, ensuring frequent failure and an absence of 'plot armor' for characters, directly countering the positivity bias prevalent in other language models.
- Context Length: Supports a 32768 token context length, allowing for extended and complex narrative arcs.
Recommended Use Cases
- Challenging Text Adventures: Ideal for generating text-based games where players can face genuine opposition, fail, and experience consequences.
- Unforgiving Role-Play: Suited for role-playing scenarios that require tension, conflict, and a realistic portrayal of danger.
- Countering Positivity Bias: A strong choice for developers seeking an AI that does not shy away from negative outcomes or difficult situations in narrative generation.
Limitations
- Second-Person Present Tense: Primarily trained on second-person present tense data, other narrative styles may yield suboptimal results.
- Single-Turn Chat Data: Trained exclusively on single-turn chat data.