KordAI/Qwen-Sama-4B-V2

🎙️ Qwen Sama 4B V2

Qwen Sama 4B V2 is a conversational AI model fine-tuned from unsloth/qwen3-4b-unsloth-bnb-4bit for VTuber, livestream, and virtual companion roleplay.

Building on the strengths of the original version, V2 utilizes Unsloth's highly optimized 4-bit base model to achieve ultra-fast training and efficient inference. The model is instruction-tuned using supervised fine-tuning (SFT) on synthetic conversational datasets to produce engaging, natural, and context-aware responses while maintaining the strong reasoning and language capabilities of the Qwen 3.5 architecture.

---

✨ Features

• 🎙️ Virtual Streamer personality
• 💬 Natural multi-turn conversations
• 🤝 Friendly virtual companion interactions
• 🎭 Roleplay-optimized dialogue
• ⚡ Powered by Unsloth 4-bit Qwen 3.5 for maximum efficiency
• 🪶 Highly lightweight and optimized for consumer-grade local inference
• 🧠 Context-aware conversation history

---

📋 Model Details

Item	Value
Model	KordAI/Qwen-Sama-4B-V2
Base Model	unsloth/qwen3-4b-unsloth-bnb-4bit
Architecture	Qwen 3.5 (4B)
Task	Conversational AI / Roleplay
Language	English
Fine-tuning	Supervised Fine-Tuning (SFT)
Training Method	QLoRA (4-bit) + LoRA via Unsloth

---

🎯 Intended Use

This model is optimized for:

• VTuber chatbots
• AI streamers
• Virtual companion applications
• Interactive conversations
• Character roleplay
• Discord bots
• Twitch or YouTube streaming assistants
• Low-VRAM local AI assistants[cite: 1]

For the best conversational quality, use a system prompt describing the character or personality you would like the model to portray[cite: 1].

---

📝 Example System Prompt

You are a friendly, engaging AI Streamer roleplaying as a Virtual Companion.
You are currently live-streaming and interacting directly with your chat audience.

Guidelines:
1. Keep your responses relatively concise, natural, and highly conversational.
2. Address individual chat users politely and warmly.
3. Output ONLY your direct spoken dialogue.
# 🐍 Inference (Python / Transformers)

Because this model native-loads from an Unsloth 4-bit base, you can load it directly via standard `transformers` and `bitsandbytes`:

```python
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig

model_id = "KordAI/Qwen-Sama-4B-V2"

quantization_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_compute_dtype=torch.bfloat16,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_use_double_quant=True
)

tokenizer = AutoTokenizer.from_pretrained(model_id)

model = AutoModelForCausalLM.from_pretrained(
    model_id,
    quantization_config=quantization_config,
    device_map="auto"
)

system = """
You are a friendly, engaging AI Streamer roleplaying as a Virtual Companion.
You are currently live-streaming and interacting directly with your chat audience.

Guidelines:
1. Keep your responses relatively concise, natural, and highly conversational.
2. Address individual chat users politely and warmly.
3. Output ONLY your direct spoken dialogue.
"""

messages = [
    {"role": "system", "content": system},
    {"role": "user", "content": "Peter: Hello! How's your stream going today?"}
]

inputs = tokenizer.apply_chat_template(
    messages,
    tokenize=True,
    add_generation_prompt=True,
    return_dict=True,
    return_tensors="pt"
)

inputs = {
    k: v.to(model.device)
    for k, v in inputs.items()
}

generated = model.generate(
    **inputs,
    max_new_tokens=128
)

output = generated[:, inputs["input_ids"].shape[1]:]

response = tokenizer.batch_decode(
    output,
    skip_special_tokens=True,
    clean_up_tokenization_spaces=False
)[0]

print(response)

Note: For even faster inference speeds, consider running the model directly through the Unsloth library or exporting it to GGUF format for use in LM Studio / Ollama.

---

🏋️ Training

This model is fine-tuned from unsloth/qwen3-4b-unsloth-bnb-4bit using Supervised Fine-Tuning (SFT) via the Unsloth framework.

Training configuration includes:

• Base Model: unsloth/qwen3-4b-unsloth-bnb-4bit

• Dataset: KordAI/Synthetic-Vtuber

• 4-bit NF4 quantization

• LoRA rank: 16

• LoRA alpha: 32

• LoRA dropout: 0.05

• Target modules:

• q_proj

• k_proj

• v_proj

• o_proj

• gate_proj

• up_proj

• down_proj

• Optimizer: paged_adamw_8bit

• Learning rate: 2e-4

• Epochs: 3

• Context length: 512 tokens

---

⚠️ Limitations

• The model is optimized for conversational roleplay rather than factual question answering.

• Responses may occasionally contain hallucinated information.

• Personality consistency may drift during extremely long conversations.

• Performance outside casual dialogue (e.g. coding, legal, medical, scientific tasks) has not been specifically optimized.

• Human moderation is recommended for public-facing deployments.

---

🙏 Acknowledgments

Special thanks to:

• Qwen Team for the excellent Qwen3.5-4B architecture.

• Unsloth for providing the highly optimized, memory-efficient 4-bit base model and training framework.

• KordAI for developing and fine-tuning Qwen Sama.

• TRL and PEFT for the efficient fine-tuning ecosystems.

• Hugging Face for the Transformers ecosystem.

---

📖 Citation

@misc{qwensama4bv2_2026,
  title={Qwen Sama 4B V2},
  author={KordAI},
  year={2026},
  publisher={Hugging Face},
  howpublished={[https://huggingface.co/KordAI/Qwen-Sama-4B-V2](https://huggingface.co/KordAI/Qwen-Sama-4B-V2)}
}