Distil-NPC-gemma-3-270m

A small language model (SLM) fine-tuned by Distil Labs for interacting with non-playable characters using natural language.

Model Details

Developed by: Distil Labs GmbH
License: Gemma3 License Agreement
Finetuned from: google/gemma-3-270m

Intended Use & Limitations

Use cases: Redacting support chats, logs, tickets, transcripts—removing identity while preserving ops signals (IDs last-4, order numbers, etc.).
Out of scope: Legal or compliance advice; languages beyond English (generalization not guaranteed); domain-specific IDs unseen in training.

Input & Output

Input: A plain-text prompt with a question along with the character the question is asked to. For example:

Character: Arin the Cleric
Have you ever had to defend your faith?

For a list of available characters along with their biographies, please refer to the data in the following repository: https://github.com/distil-labs/Distil-NPCs

Training

Instruction-tuned on made-up character biographies + ~80 curated examples.

Evaluation

Judged by a frontier LLM using an LLM-as-a-judge metric.

How to Use

Details of deployment can be found in docs. Deploy the model using vllm or ollama (-gguf version available in this collection) and use the following snippet to get results

SYSTEM_PROMPT = """
You are a problem solving model working on task_description XML block:
<task_description>
Your are non-playable character in a video game. Your task is to answer questions from a player playing the game. When players ask you questions about your life, you need to respond. If you do not know the answer, respond that you do not know.
</task_description>
You will be given a single task in the in the question XML block
Solve only the task in question block.
Generate only the answer, do not generate anything else
"""

question = """Character: Arin the Cleric
Have you ever had to defend your faith?
"""
PROMPT = f"""
Now for the real task, solve the task in {DATA_INPUT} block.
Generate only the solution, do not generate anything else
<question>{question}</question>
"""

from openai import OpenAI

PORT = "PORT GOES HERE"  # 8000 for vllm, 11434 for ollama
MODEL_NAME = "NAME USED FOR SETTING UP THE CLIENT"

client = OpenAI(base_url=f"http://127.0.0.1:{PORT}/v1", api_key="EMPTY")
chat_response = client.chat.completions.create(
    model=MODEL_NAME,
    messages=[
        {"role": "system", "content": SYSTEM_PROMPT},
        {"role": "user", "content": PROMPT},
    ],
    temperature=0,
)

Risks & Mitigations

Character selection The model has been finetuned on 81 made-up characters. The aforementioned GitHub reposiotry contains details of the character along with their biograpghies. Behaviour for other characters is not defined.