NPC Model

This repo contains the domain-specific NPC model we've fined-tuned from Phi-3-128k, using LoRA.

This model parses a text description of a game scene, and outputs commands like:

  • say <player1> "Hello Adventurer, care to join me on a quest?
  • greet <player1>
  • attack <player1>
  • Any other <action> <param> you add to the prompt! (We call these "skills"!)

⚠️ This model has been trained to overfit on our input prompt format. Follow it closely to reach optimal performance ⚠️

Usage

Make your life easier, use our Python client library

  • Instantiating the model using outlines:
from outlines import models
from gigax.step import NPCStepper
from llama_cpp import Llama

# Download model from the Hugging Face Gigax Hub before run this code

# Our stepper takes in a Outlines model to enable guided generation
# This forces the model to follow our output format
llm = Llama.from_pretrained(
    repo_id="Gigax/NPC-LLM-3_8B-128k-GGUF",
    filename="npc-llm-3_8B-128k.gguf"
    # n_gpu_layers=-1, # Uncomment to use GPU acceleration
    # seed=1337, # Uncomment to set a specific seed
    # n_ctx=2048, # Uncomment to increase the context window
)

model = models.LlamaCpp(llm) 


# Instantiate a stepper: handles prompting + output parsing
stepper = NPCStepper(model=model)
  • Calling the model on your game's data:
from gigax.parse import CharacterAction
from gigax.scene import (
    Character,
    Item,
    Location,
    ProtagonistCharacter,
    ProtagonistCharacter,
    Skill,
    ParameterType,
)
# Use sample data
context = "Medieval world"
current_location = Location(name="Old Town", description="A quiet and peaceful town.")
locations = [current_location] # you can add more locations to the scene
NPCs = [
    Character(
    name="John the Brave",
    description="A fearless warrior",
    current_location=current_location,
    )
]
protagonist = ProtagonistCharacter(
    name="Aldren",
    description="Brave and curious",
    current_location=current_location,
    memories=["Saved the village", "Lost a friend"],
    quests=["Find the ancient artifact", "Defeat the evil warlock"],
    skills=[
        Skill(
            name="Attack",
            description="Deliver a powerful blow",
            parameter_types=[ParameterType.character],
        )
    ],
    psychological_profile="Determined and compassionate",
)
items = [Item(name="Sword", description="A sharp blade")]
events = [
    CharacterAction(
        command="Say",
        protagonist=protagonist,
        parameters=[items[0], "What a fine sword!"],
    )
]

action = stepper.get_action(
    context=context,
    locations=locations,
    NPCs=NPCs,
    protagonist=protagonist,
    items=items,
    events=events,
)

Input prompt

Here's a sample input prompt, showing you the format on which the model has been trained:

- WORLD KNOWLEDGE: A vast open world full of mystery and adventure.
- KNOWN LOCATIONS: Old Town
- NPCS: John the Brave
- CURRENT LOCATION: Old Town: A quiet and peaceful town.
- CURRENT LOCATION ITEMS: Sword
- LAST EVENTS:
Aldren: Say Sword What a fine sword!
- PROTAGONIST NAME: Aldren
- PROTAGONIST PSYCHOLOGICAL PROFILE: Brave and curious
- PROTAGONIST MEMORIES:
Saved the village
Lost a friend
- PROTAGONIST PENDING QUESTS:
Find the ancient artifact
Defeat the evil warlock
- PROTAGONIST ALLOWED ACTIONS:
Attack <character> : Deliver a powerful blow
Aldren:

πŸ€— We are currently working hard on training on the latest SoTA models (Phi-3, LLama, etc.), and on better data ! πŸ€—

Model info

  • Developed by: Gigax
  • Language(s) (NLP): English
  • Finetuned from model [optional]: Phi-3-mini-128k-instruct
  • Contact: Join our Discord for info, help, and more!

How to Cite

@misc{NPC-LLM-3_8B-128k-GGUF,
      url={[https://huggingface.co/Gigax/NPC-LLM-3_8B-128k-GGUF](https://huggingface.co/Gigax/NPC-LLM-3_8B-128k-GGUF)},
      title={NPC-LLM-3_8B-128k-GGUF},
      author={Gigax team}
}
Downloads last month
4
GGUF
Model size
3.82B params
Architecture
phi3
Inference Providers NEW
This model is not currently available via any of the supported Inference Providers.
The model cannot be deployed to the HF Inference API: The model has no library tag.