Wayfarer-Large-70B-Llama-3.3
| Property | Value |
|---|---|
| Parameter Count | 70B |
| Base Model | Llama 3.3 70B Instruct |
| Author | LatitudeGames |
| Model URL | https://huggingface.co/LatitudeGames/Wayfarer-Large-70B-Llama-3.3 |
What is Wayfarer-Large-70B-Llama-3.3?
Wayfarer-Large-70B-Llama-3.3 is a specialized language model designed to address a common criticism in AI storytelling: the tendency of models to be overly positive and protective of characters. Built on top of Llama 3.3 70B Instruct, this model introduces a unique approach to narrative AI by embracing conflict, tension, and consequences in storytelling.
Implementation Details
The model was trained using a carefully curated mixture of data: 33% text adventure data (8k tokens), 33% roleplay data (4k tokens), and 33% SlimOrca Sonnet subset. The training process involved simulating playthroughs of published character creator scenarios, with five distinct user archetypes generating unique samples. The model uses second-person present tense as its primary narrative style and follows the Llama 3 prompt format.
- Training incorporated blind simulation between narrator and user models
- Optimized for 8k token length or until character death
- Implements a baseline temperature of 0.8 and repetition penalty of 1.05
- Utilizes min_p setting of 0.05
Core Capabilities
- Generation of challenging and dangerous narrative experiences
- Creation of stories with real stakes and consequences
- Support for multiple narrative perspectives despite second-person training
- Advanced conflict and tension management in storytelling
Frequently Asked Questions
Q: What makes this model unique?
This model breaks away from the conventional "nice AI" paradigm by specifically training for pessimistic outcomes and genuine conflict, making it ideal for gaming and adventure scenarios where challenge and risk are desired.
Q: What are the recommended use cases?
The model excels in text-based adventure games, role-playing scenarios, and interactive fiction where genuine challenge and consequences are desired. It's particularly suited for platforms like AI Dungeon where engaging narrative experiences are priority.
