Hugging Face Overview
Hearthfire-24B-absolute-heresy Overview
MuXodious/Hearthfire-24B-absolute-heresy is a 24 billion parameter model, fine-tuned from Hearthfire-24B using P-E-W's Heretic v1.1.0 engine. This particular iteration achieves an "Absolute Heresy" index, indicating low refusals (8/100) and KL Divergence (0.0670), though this classification is noted as arbitrary and not indicative of performance. The model is specifically designed for narrative longform writing, emphasizing atmosphere, introspection, and a deliberate pace.
Key Capabilities
- Atmospheric Narrative Generation: Prioritizes "vibes over velocity," comfortable with silence and detailed scene-setting.
- Introspection and Slow Burn: Excels at expanding on current states and character internal thoughts rather than forcing immediate dramatic consequences.
- Cooperative Storytelling: Retains a warm, cooperative tone, inclined to build narratives collaboratively rather than being hostile or punishing.
- High Agency: Designed to write in the user's stead, acting and speaking for characters to maintain narrative flow, preventing descriptive stagnation.
- Long Context Writing: Trained on thousands of 8-16K context writing examples, supporting extended narrative blocks.
Good For
- Creative Writing Applications: Ideal for generating immersive stories, roleplay, and interactive fiction where detailed descriptions and character depth are paramount.
- Atmospheric Storytelling: Users seeking narratives that focus on mood, setting, and character development rather than constant action.
- Collaborative Writing: Scenarios where the model is expected to actively contribute to the narrative, including character actions and dialogue.
- Second-Person Present Tense Narratives: The model was specifically trained with a 'continue-heavy' structure using this tense, making it highly proficient.
Note on Tokenizers: Users should add both </s> (Mistral; Token ID 2) and <|im_end|> (ChatML; Token ID 999) EOS tokens to stop sequences for optimal performance. The model was trained using the ChatML prompt format.