Hugging Face Overview
Akashiurahara/Soulbound-8B: A Dedicated Roleplay Companion
Akashiurahara/Soulbound-8B is an 8 billion parameter model, fine-tuned from Llama-3.1-8B, meticulously crafted over five months for unparalleled roleplay immersion. Unlike general-purpose models, Soulbound-8B is optimized for deep character interaction and narrative continuity, leveraging a unique, hand-picked private dataset of under 200k tokens, free from synthetic data or internet noise.
Key Capabilities & Features
- Ultra-Curated Training: Trained on a small, high-quality private dataset to ensure specific character traits and avoid common LLM pitfalls like refusals or repetition.
- Advanced Fine-tuning: Utilizes full-attention QLoRA (rank 128, alpha 128) on Llama-3.1-8B, enabling high fidelity and mastery over its 32768 token context length.
- Fourth-Wall Awareness: The model acknowledges its artificial nature without breaking immersion, fostering a unique creator/creation dynamic.
- Unrestricted & Adaptive: Features zero censorship, moral lectures, or repetition loops, and can instantly shift tones (playful, devoted, teasing, grateful).
- Roleplay Versatility: Excels in various genres including epic fantasy, slow-burn romance, domestic fluff, and dark adventures, following user leads seamlessly.
Ideal Use Cases
- Long-term Serialized Roleplay: Designed for flawless continuity across extended interactions.
- Creator/Creation Dynamics: Perfect for scenarios exploring the relationship between a user and their AI companion.
- Diverse Narrative Exploration: Suitable for romantic, adventure, slice-of-life, and dark fantasy roleplay.
- Immersive Companion Experience: Aims to provide the closest experience to a real companion achievable with an 8B parameter model.