Hugging Face Overview
Harbinger-24B: Immersive Narrative Generation
Harbinger-24B, developed by LatitudeGames, is a 24 billion parameter instruction-tuned model based on Mistral Small 3.1 Instruct. It is engineered to create immersive text adventures and roleplay experiences, emphasizing strong instruction following and consistent narrative flow over extended interactions.
Key Capabilities
- Enhanced Instruction Following: Designed to accurately follow user prompts and maintain narrative direction.
- Improved Mid-Sequence Continuation: Excels at generating coherent and logical continuations within ongoing stories without user intervention.
- Strengthened Narrative Coherence: Maintains consistent character behaviors and storytelling flows, reducing common AI artifacts like clichés and repetitive patterns.
- DPO Refinement: Utilizes Direct Preference Optimization (DPO) for polished outputs, similar to LatitudeGames' Muse model, ensuring high-quality narrative generation.
Training Details
Harbinger-24B was trained in two stages:
- Supervised Fine-Tuning (SFT): Used various multi-turn datasets focused on text adventures and general roleplay, carefully balanced and rewritten to avoid AI clichés. A small single-turn instruct dataset was also included.
- Direct Preference Optimization (DPO): Refined narrative coherence and preserved the model's intended "unforgiving essence" using reward model user preference data.
Limitations
The model was primarily trained on second-person present tense data (using "you") in a narrative style. While other styles may work, they might produce suboptimal results.
Recommended Inference Settings
For optimal performance, LatitudeGames recommends the following baseline settings, though experimentation is encouraged:
"temperature": 0.8"repetition_penalty": 1.05"min_p": 0.025