Qwen2.5-7B-Instruct-SFT-FT1

A LoRA fine-tuned version of Qwen/Qwen2.5-7B-Instruct for creative advertising concept generation.

Model Description

This model was fine-tuned using Supervised Fine-Tuning (SFT) with LoRA adapters on a curated dataset of creative advertising campaigns. The model specializes in generating unconventional, attention-grabbing advertising ideas including experiential marketing, guerrilla campaigns, memorable mascots, and viral concepts.

Training Objective

Transform the base Qwen2.5-7B-Instruct model into a creative advertising expert capable of generating innovative campaign ideas that go beyond traditional advertising formats.

Training Results

Metric	Value
Average Train Loss	1.219
Final Train Loss	0.972
Eval Loss	0.988
Train/Eval Token Accuracy	78.75%
Total Training Steps	155
Training Runtime	379 seconds (~6.3 min)
Gradient Norm	0.44

Training Curves

View detailed training metrics on Weights & Biases: W&B Training Run

Analysis of Results

• No Overfitting: Train loss (0.972) and eval loss (0.988) are nearly identical with a gap of only 0.016
• Stable Training: Gradient norm remained healthy at 0.44 throughout training
• Good Generalization: The close train/eval losses indicate the model learned generalizable patterns rather than memorizing training data
• Task-Appropriate Loss: The loss values (~0.97-1.2) reflect the inherently open-ended nature of creative writing tasks, where multiple valid outputs exist for each input

Training Configuration

LoRA Parameters

Parameter	Value
LoRA Rank (r)	32
LoRA Alpha	64
LoRA Dropout	0.1
Target Modules	q_proj, k_proj, v_proj, o_proj
Bias	none
Task Type	CAUSAL_LM

Training Hyperparameters

Parameter	Value
Learning Rate	1e-4
LR Scheduler	Cosine
Warmup Ratio	0.1
Epochs	5
Batch Size (per device)	2
Gradient Accumulation	8
Effective Batch Size	16
Max Sequence Length	2048
Optimizer	AdamW
Weight Decay	0.01
Precision	bf16

Dataset

• Dataset: 55mvresearch/sft-v1-singleturn-ads-creativity
• Total Examples: 539
• Train Split: 485 examples (90%)
• Eval Split: 54 examples (10%)

System Prompt

The model was trained with the following system prompt:

You are an award-winning creative director at a top advertising agency, known for developing unconventional campaigns that capture attention and become cultural moments.

Your approach to creativity:
- Think beyond traditional advertising formats - consider experiential, guerrilla, interactive, and unexpected mediums
- Create ideas that are inherently shareable and generate organic buzz
- Develop memorable mascots, characters, or visual hooks that become brand assets
- Find the unexpected angle that makes people stop, look, and remember
- Transform ordinary products into extraordinary stories

When crafting campaign ideas:
- Lead with the big creative concept, not the product features
- Describe the execution vividly so the idea comes to life
- Consider how the idea extends across touchpoints and generates earned media
- Make the brand the hero through clever, surprising creative choices

Output format: Write a single, flowing long-form description of the campaign idea. Do not use titles, headings, bullet points, or numbered lists. Paint the picture of the idea as a continuous narrative.

Usage

Installation

pip install transformers peft torch accelerate

Loading the Model

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

# Load base model
base_model = "Qwen/Qwen2.5-7B-Instruct"
model = AutoModelForCausalLM.from_pretrained(
    base_model,
    torch_dtype="auto",
    device_map="auto",
)

# Load LoRA adapter
model = PeftModel.from_pretrained(model, "55mvresearch/Qwen2.5-7B-Instruct-SFT-FT1")

# Load tokenizer
tokenizer = AutoTokenizer.from_pretrained(base_model)

Inference

system_prompt = """You are an award-winning creative director at a top advertising agency, known for developing unconventional campaigns that capture attention and become cultural moments.

Your approach to creativity:
- Think beyond traditional advertising formats - consider experiential, guerrilla, interactive, and unexpected mediums
- Create ideas that are inherently shareable and generate organic buzz
- Develop memorable mascots, characters, or visual hooks that become brand assets
- Find the unexpected angle that makes people stop, look, and remember
- Transform ordinary products into extraordinary stories

When crafting campaign ideas:
- Lead with the big creative concept, not the product features
- Describe the execution vividly so the idea comes to life
- Consider how the idea extends across touchpoints and generates earned media
- Make the brand the hero through clever, surprising creative choices

Output format: Write a single, flowing long-form description of the campaign idea. Do not use titles, headings, bullet points, or numbered lists. Paint the picture of the idea as a continuous narrative."""

user_prompt = "Create a quirky campaign for a new energy drink targeting Gen Z gamers."

messages = [
    {"role": "system", "content": system_prompt},
    {"role": "user", "content": user_prompt}
]

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

model_inputs = tokenizer([text], return_tensors="pt").to(model.device)

generated_ids = model.generate(
    **model_inputs,
    max_new_tokens=512,
    temperature=0.7,
    top_p=0.9,
    do_sample=True,
)

generated_ids = [
    output_ids[len(input_ids):]
    for input_ids, output_ids in zip(model_inputs.input_ids, generated_ids)
]

response = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
print(response)

Training Infrastructure

• Hardware: NVIDIA A100 (80GB)
• Training Framework: TRL (Transformer Reinforcement Learning)
• PEFT Library: Hugging Face PEFT
• Experiment Tracking: Weights & Biases

Limitations

• Trained on a relatively small dataset (539 examples)
• Specialized for creative advertising - may not perform as well on other tasks
• Generates ideas in English only
• Creative output quality may vary and should be reviewed by humans

License

This model is released under the Apache 2.0 license, following the base model's license.

Citation

If you use this model, please cite:

@misc{qwen25-7b-ads-creativity,
  author = {55mvresearch},
  title = {Qwen2.5-7B-Instruct-SFT-FT1: Creative Advertising LoRA},
  year = {2025},
  publisher = {Hugging Face},
  url = {https://huggingface.co/55mvresearch/Qwen2.5-7B-Instruct-SFT-FT1}
}

Acknowledgments

• Qwen Team for the base model
• Hugging Face for the training infrastructure and libraries