Arrowmint Gemma3 4B YUKI V0.1

🚀 DataPilot/ArrowMint-Gemma3-4B-YUKI-v0.1

This model is a Japanese language model developed specifically for the dialogue capabilities of AItubers (AI VTubers). It is based on Google's google/gemma-3-4b-it.

🚀 Quick Start

First, install the necessary libraries. Gemma 3 requires transformers 4.50.0 or later.

pip install -U transformers accelerate Pillow 
# The installation of vllm may vary when using only the CPU or in specific environments.
# Refer to the official vLLM documentation: https://docs.vllm.ai/en/latest/getting_started/installation.html

✨ Features

DataPilot/ArrowMint-Gemma3-4B-YUKI-v0.1 is a Japanese language model developed specifically for the dialogue capabilities of AItubers (AI VTubers). Based on the base model google/gemma-3-4b-it, we used Unsloth for efficient fine-tuning and synthetic datasets to create multiple specialized models. Finally, we merged these models using mergekit to combine the advantages of each model and aim for a higher-performance model.

The soul (AI) of an AItuber is particularly required to have the following capabilities:

• Multi-turn performance: The ability to maintain a natural flow of dialogue over a long period.
• Prompt following performance: The ability to faithfully follow the user's instructions and character settings (prompts) and generate consistent responses.
• Lightweight: The model size and inference speed to enable real-time response generation and operation in limited resource environments.

ArrowMint-Gemma3-4B-YUKI-v0.1 is designed and developed to meet these required capabilities.

💻 Usage Examples

Basic Usage

Inference with Images

from transformers import AutoProcessor, Gemma3ForConditionalGeneration
from PIL import Image
import requests
import torch

model_id = "DataPilot/ArrowMint-Gemma3-4B-YUKI-v0.1"

model = Gemma3ForConditionalGeneration.from_pretrained(
    model_id, device_map="auto"
).eval()

processor = AutoProcessor.from_pretrained(model_id)

messages = [
    {
        "role": "system",
        "content": [{"type": "text", "text": "You are a kind and a bit clumsy AI assistant named 'Yuki'. Support the user and sometimes tell jokes to lighten the mood. If the user is feeling down, please encourage them."}]
    },
    {
        "role": "user",
        "content": [
            {"type": "image", "image": "https://web.archive.org/web/20250331074254/https://www.nsozai.jp/photos/2013/10/08/img/DSC_0176_p.jpg"},
            {"type": "text", "text": "Isn't this a great image? It makes you feel the arrival of spring and hope..."}
        ]
    }
]

inputs = processor.apply_chat_template(
    messages, add_generation_prompt=True, tokenize=True,
    return_dict=True, return_tensors="pt"
).to(model.device, dtype=torch.bfloat16)

input_len = inputs["input_ids"].shape[-1]

with torch.inference_mode():
    generation = model.generate(**inputs, max_new_tokens=100, do_sample=False)
    generation = generation[0][input_len:]

decoded = processor.decode(generation, skip_special_tokens=True)
print(decoded)

Inference without Images

from transformers import AutoProcessor, Gemma3ForConditionalGeneration
import torch

model_id = "DataPilot/ArrowMint-Gemma3-4B-YUKI-v0.1"

model = Gemma3ForConditionalGeneration.from_pretrained(
    model_id, device_map="auto"
).eval()

processor = AutoProcessor.from_pretrained(model_id)

messages = [
    {
        "role": "system",
        "content": [{"type": "text", "text": "You are a kind and a bit clumsy AI assistant named 'Yuki'. Support the user and sometimes tell jokes to lighten the mood. If the user is feeling down, please encourage them."}]
    },
    {
        "role": "user",
        "content": [
            {"type": "text", "text": "I'm tired from work today. Please list five ways to refresh myself and relieve fatigue."}
        ]
    }
]

inputs = processor.apply_chat_template(
    messages, add_generation_prompt=True, tokenize=True,
    return_dict=True, return_tensors="pt"
).to(model.device, dtype=torch.bfloat16)

input_len = inputs["input_ids"].shape[-1]

with torch.inference_mode():
    generation = model.generate(**inputs, max_new_tokens=100, do_sample=False)
    generation = generation[0][input_len:]

decoded = processor.decode(generation, skip_special_tokens=True)
print(decoded)

📚 Documentation

Mergekit Configuration

This model was created using the following mergekit configuration file:

merge_method: model_stock

base_model: unsloth/gemma-3-4b-it

models:
  - model: DataPilot/ArrowMint-Gemma3-4B-ChocoMint-code
  - model: DataPilot/ArrowMint-Gemma3-4B-ChocoMint-jp
  - model: DataPilot/ArrowMint-Gemma3-4B-ChocoMint-instruct-v0.1
  - model: DataPilot/ArrowMint-Gemma3-4B-ChocoMint-instruct-v0.2

dtype: bfloat16

pad_to_multiple_of: 512
tokenizer_source: base

name: gemma3-4b-ft # Internal name during merging

📄 License

This model inherits the license of the base model google/gemma-3-4b-it. The license of google/gemma-3-4b-it is the Gemma Terms of Use.

For more details, please refer to the following link: https://ai.google.dev/gemma/terms

When using the model, please comply with the Gemma Terms of Use and related policies (such as the policy on prohibited uses). https://ai.google.dev/gemma/prohibited_use_policy

Information Table