Kavyaah/copywriting-llm

🪄 Copywriting LLM

Generate short, high-converting push notifications and ad copies.

This model is fine-tuned on curated marketing and app-notification data using Mistral-7B-Instruct (Unsloth) with LoRA and 4-bit quantization. It creates concise, catchy lines for offers, FOMO alerts, food cravings, re-engagement, and festive campaigns.

Model Details

Property Value Base Model unsloth/mistral-7b-instruct-v0.3 Fine-Tuning LoRA (r = 16, α = 16, dropout = 0.0) Quantization 4-bit (QLoRA NF4) Dataset 3 000 handcrafted marketing prompts & responses Task Causal Language Modeling for short-form copywriting Context Length 2048 tokens

Usage

#
from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

# Load tokenizer & model
tokenizer = AutoTokenizer.from_pretrained("Kavyaah/copywriting-llm")
model = AutoModelForCausalLM.from_pretrained("Kavyaah/copywriting-llm", torch_dtype="auto")
model.eval()

# Function to generate push notification
def generate_copy(brand, offer, tone="fun", max_new_tokens=40):
    prompt = f"""You are an expert marketing copywriter.
Write a short, catchy push notification in a {tone} tone.
It should promote {brand}'s offer: "{offer}".
Keep it under 20 words, engaging, and persuasive."""
    
    inputs = tokenizer(prompt, return_tensors="pt")
    with torch.no_grad():
        outputs = model.generate(
            **inputs,
            max_new_tokens=max_new_tokens,
            temperature=0.9,
            top_p=0.9,
            do_sample=True
        )
    return tokenizer.decode(outputs[0], skip_special_tokens=True)

Example

print(generate_copy("Zomato", "Flat 60% off on dinner combos this weekend!"))
#
Example Output
Dinner’s calling 🍽️ 60% off on Zomato combos—grab your feast before the weekend ends!

Evaluation

Metric Result

Human rated copy quality 8.5 / 10

Tone accuracy (fun & playful) 93 %

Avg token length 18 words

Intended Use

Generating push notifications, app banners, and micro-ad copies

Creative assistants for marketing and growth teams

Automating A/B test copy variants for offers and sales

Limitations

May produce overly playful or repetitive content if prompts are vague

Trained only for short-form marketing copywriting

Avoid using for sensitive topics or regulated industries

Technical Configuration

Parameter Value

Optimizer AdamW (8-bit)

Learning Rate 2 × 10⁻⁴

Epochs 2

Gradient Accumulation 4

Batch Size (effective) 8

Quantization 4-bit QLoRA

Training Data Categories

Category Example

Sale / Offer “Diwali deals up to 50% off ✨”

Food Craving “Lunch o’clock alert! Your cravings just went live 🍛”

FOMO “Blink and it’s gone 👀 Flash sale ends in 2 hours!”

Re-engagement “We miss your clicks 😢 Come back for something tasty!”

Festive “Play with colors, not your budget! Holi offers just dropped 🎨”

Fashion “New drops just landed 💃 Make your wardrobe jealous!”

License

MIT License - open for research and non-commercial use.

Please credit Kavyaa / Copywriting LLM if you use this model in public projects.

Acknowledgements

Fine-tuned using Unsloth for 2× faster training

Base weights from Mistral-7B-Instruct v0.3

Created by Kavyaa for creative and marketing AI research