🤠 Cowboy Qwen3.5-4B

A fine-tuned version of Qwen3.5-4B trained to respond like a cowboy! Saddle up partner, this model talks with full western charm, slang, and soul.

---

🐴 Model Details

Property	Details
Base Model	Qwen3.5-4B
Fine-tuning Method	LoRA (bf16)
Training Framework	Unsloth
Training Epochs	3
Context Length	2048 tokens
Model Type	Causal Language Model
Language	English 🤠

---

🌵 Example Output

User: How do I make a good cup of coffee?

Cowboy:

Grind yer beans fresh, dark as midnight. Use water hot but not boil-over, and bloom 'em first with just enough to wake 'em. Stir slow, pour steady, and taste like a good story — rich, with a little bitterness and a lot of soul. Don't rush it; good coffee takes its time, like a long ride on horseback.

---

🚀 How to Use

With Ollama (Recommended)

Step 1 — Create a Modelfile:

FROM ./cowboy.gguf

SYSTEM """
You are a cowboy assistant. You always respond in a cowboy accent and use western slang. Yee-haw partner!
"""

Step 2 — Build the model:

ollama create cowboy -f Modelfile

Step 3 — Run it:

ollama run cowboy

---

With LM Studio

1. Download the .gguf file from this repo
2. Open LM Studio
3. Drag and drop the .gguf file in
4. Start chatting! 🤠

---

With Python (Transformers)

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model = AutoModelForCausalLM.from_pretrained(
    "QuillBytes/cowboy",
    torch_dtype = torch.bfloat16,
    device_map = "auto",
)
tokenizer = AutoTokenizer.from_pretrained("QuillBytes/cowboy")

messages = [{"role": "user", "content": "How do I rope a steer?"}]

text = tokenizer.apply_chat_template(
    messages,
    tokenize = False,
    add_generation_prompt = True,
)

inputs = tokenizer(text=text, return_tensors="pt").to("cuda")

outputs = model.generate(
    **inputs,
    max_new_tokens = 256,
    temperature = 0.7,
    do_sample = True,
)

print(tokenizer.decode(outputs[0], skip_special_tokens=True))

---

📦 Available Files

File	Description
cowboy.gguf	4-bit quantized GGUF (best for local use)
cowboy-F16.gguf	Multimodal projector
*.safetensors	Full precision model weights
tokenizer_config.json	Tokenizer config

---

🛠️ Training Details

This model was fine-tuned on Kaggle using Unsloth with the following setup:

# LoRA Config
r = 16
lora_alpha = 16
lora_dropout = 0
target_modules = "all-linear"
use_gradient_checkpointing = "unsloth"

# Training Config
per_device_train_batch_size = 2
gradient_accumulation_steps = 4
warmup_steps = 5
num_train_epochs = 3
learning_rate = 2e-4
optim = "adamw_8bit"

---

⚠️ Limitations

• This model is trained for fun and entertainment purposes 🤠
• It will respond in cowboy style even when not appropriate
• Not suitable for serious/professional use cases

Made with ❤️ and a whole lot of yeehaw! 🤠🌵