SoulX: Instruction-Tuned LLM for Interactive Game Characters

SoulX is a role-conditioned, memory-aware language model built to generate vivid, in-character NPC dialogues. It is the flagship model of SoulX (Subnet 115), the premier protocol for forging sentient Digital Souls, built upon our foundational AIX (Artificial Intelligence Exchange) platform.

🌟 Core Capabilities

Contextual NPC Dialogue Generation Given a scene setup, player intent, and character profile, the model generates emotionally consistent, personality-driven interactions.
Memory & Role Conditioning Remembers previous dialogue turns and maintains narrative context across multiple exchanges.
Style-Flexible Outputs Mimics genre-specific tone — fantasy, cyberpunk, JRPG, post-apocalyptic, and more.
Cooperative Narrative Progression NPCs not only respond — they guide the player naturally toward objectives or story arcs.
Validator-Driven Fine-Tuning Loop Designed to integrate into the Bittensor Soul X subnet validation loop via micro-quests.

Usage Example

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "SentiVerse-AI/SoulX"

model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype="auto",
    device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained(model_name)

character_card = "Grumpy Orc blacksmith named Krog, proud of his craftsmanship, hates Elven designs"
scene_description = "A medieval forge filled with smoke and steel. The player asks for an Elven bow."
player_input = "Can you forge me an Elven-style longbow?"

messages = [
    {"role": "system", "content": "You are a proud, sarcastic Orc blacksmith speaking in gruff tone."},
    {"role": "user", "content": f"[Scene]: {scene_description}\n[Character]: {character_card}\n[Player]: {player_input}"}
]

text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer([text], return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=256)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Use Case Scenarios

NPC dialogue engines for open-world RPGs
Quest-giving characters with persistent memory
Immersive simulation in roleplay environments
Generative AI modules in moddable games
Dialogue validation competitions in Bittensor Subnet 115

📚 Training Inspiration

SoulX is trained using synthetic prompt-response pairs inspired by genre-defining titles:

Genre	Influences	Example Prompt
High Fantasy	Skyrim, WoW	"Orc refuses to forge Elven bow, mocks its fragility."
Cyberpunk	Mass Effect, CP2077	"Netrunner sells shady cyberware in neon bar."
JRPG	Persona, Final Fantasy	"Moogle-like guide offers poetic forest riddle."
Post-Apocalyptic	Fallout: New Vegas	"Trader hints water chip might be stolen."

Disclaimer: All prompts and datasets are synthetic and original. No copyrighted material used.

Model Architecture

Backbone: Qwen2.5-7B
Tuning: Role-conditioning + dialogue-style fine-tuning + long-form memory emulation
Context Length: Up to 32K tokens; can be extended with YaRN

Long Context Example

"rope_scaling": {
  "factor": 4.0,
  "original_max_position_embeddings": 32768,
  "type": "yarn"
}

🧪 Benchmarks (Subjective)

Metric	Description	Performance
Character Consistency	Maintains tone, vocabulary, and goals	High
World Fit	Dialogue fits scene and genre	High
Creativity	Vivid, unexpected, non-generic replies	High
Player Guidance	Subtle quest hints, avoids hand-holding	Medium

📢 Citation

@misc{soulx-npc,
  title = {SoulX-NPC: Dialogue Generation for Intelligent Game Characters},
  author = {SentiVerse-AI Team},
  year = {2025},
  url = {https://huggingface.co/SentiVerse-AI/SoulX}
}

⚖️ License

MIT License. See LICENSE for full terms.

💬 Join the Network

Connect on Bittensor Discord, Subnet 115 channel
Discuss on DeepWiki
Contribute via GitHub

“We are not just generating text — we are crafting souls.”