Qwen3-4B-ORPO-Advertisement-GGUF

GGUF quantized versions of Qwen3-4B-ORPO-Advertisement, a specialized language model for Advertisement Strategy, Advertisement Analysis, and Performance Marketing.

Model Overview

Qwen3-4B-ORPO-Advertisement-GGUF provides quantized GGUF versions of the fine-tuned yothinS/Qwen3-4B-ORPO-Advertisement model.

The original model is based on Qwen/Qwen3-4B and fine-tuned using Odds Ratio Preference Optimization (ORPO) with the Advertisement Specialist Dataset.

This repository is intended to make the model easier to run locally using GGUF-compatible inference engines, with different quantization levels to balance memory usage, model size, inference speed, and output quality.

Key Features

• Advertisement strategy and campaign planning
• Advertising performance analysis
• Marketing KPI interpretation, including CTR, CPC, CVR, CPA, and ROAS
• Audience targeting and segmentation
• Campaign optimization and budget allocation support
• Creative strategy, advertising copy, headlines, and calls to action
• Conversion optimization and advertising funnel analysis
• Thai and English marketing assistance, with performance depending on the task

These are intended capabilities, not guarantees of accuracy.

Available Quantizations

The repository provides the following quantization variants.

Quantization quality, memory requirements, and inference speed depend on the hardware, inference engine, context length, and workload. Actual GGUF file sizes may vary.

Recommended starting point: Q4_K_M offers a practical balance for many local inference setups. If memory permits and preserving output quality is a priority, consider Q5_K_M or Q6_K.

Model Details

Intended Use

This model is designed for:

• Marketing professionals and advertising specialists
• Performance marketers and campaign analysts
• Business owners and digital marketing teams
• Developers building local marketing assistants
• Researchers exploring domain-specific language model fine-tuning and quantization

Potential applications include analyzing advertising KPIs, brainstorming creative concepts, evaluating campaign performance, and generating data-informed marketing recommendations.

The model should not be treated as a substitute for verified campaign data, controlled experiments, or professional judgment.

How to Use

Option 1: Run with llama.cpp

Install or build a recent version of llama.cpp that supports the model architecture and quantization used in this repository.

Download a GGUF file from this repository and run:

llama-cli \
  -m ./Qwen3-4B-ORPO-Advertisement-Q4_K_M.gguf \
  -cnv \
  -n 512

Replace the filename with the exact GGUF file you downloaded.

For interactive conversations, use a compatible chat template. The model's tokenizer and chat-template metadata should be preserved during GGUF conversion.

Option 2: Run with LM Studio

1. Open LM Studio.
2. Find or import the downloaded GGUF file.
3. Load the model using a compatible runtime.
4. Configure context length and GPU offloading according to your available memory.
5. Start a conversation about advertising strategy or campaign analysis.

Compatibility depends on the LM Studio version and its underlying inference runtime.

Option 3: Run with Hugging Face GGUF Support

You can use a GGUF file with compatible Transformers versions for supported inference workflows.

from transformers import AutoTokenizer, AutoModelForCausalLM

model_id = "yothinS/Qwen3-4B-ORPO-Advertisement"

tokenizer = AutoTokenizer.from_pretrained(model_id)

model = AutoModelForCausalLM.from_pretrained(
    model_id,
    gguf_file="Qwen3-4B-ORPO-Advertisement-Q4_K_M.gguf",
    device_map="cpu",
)

messages = [
    {
        "role": "user",
        "content": (
            "An advertising campaign spent 10000 THB "
            "and generated 40000 THB in attributed revenue. "
            "Calculate ROAS and explain its limitations."
        ),
    }
]

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

inputs = tokenizer(prompt, return_tensors="pt")

outputs = model.generate(
    **inputs,
    max_new_tokens=256,
    do_sample=False,
)

response = tokenizer.decode(
    outputs[0][inputs["input_ids"].shape[1]:],
    skip_special_tokens=True,
)

print(response)

Note: GGUF support through Transformers depends on the installed version and model architecture. This is an alternative workflow, not the preferred way to run GGUF. For local GGUF inference, llama.cpp is generally the more direct option. The original Hugging Face model repository must provide the tokenizer files, and enable_thinking must be supported by its chat template.

Quantization Notes

Quantization reduces the precision used to store model weights, typically reducing disk space and memory requirements.

• Q2_K: Prioritizes a smaller model footprint and may cause a greater loss of response quality.
• Q3_K_M: A compromise for systems with tighter memory constraints.
• Q4_K_M: A practical balance of size and output quality.
• Q5_K_M: Uses more storage to preserve additional precision.
• Q6_K: A higher-precision quantized option.
• Q8_0: An 8-bit quantized option that generally preserves more precision than lower-bit variants.
• F16: Stores weights in 16-bit floating-point format without low-bit weight quantization.

Quantization does not improve the model's underlying knowledge. Actual quality differences should be established through evaluation rather than assumed.

Evaluation

The GGUF variants should be evaluated against the original fine-tuned model using the same prompts and evaluation procedure.

Recommended benchmarks include:

• Advertising KPI calculation accuracy
• Campaign analysis quality
• Marketing strategy relevance
• Instruction-following performance
• Thai and English response quality
• Inference speed and memory usage

For a fair comparison, use the same held-out test set, comparable context settings, and consistent generation parameters.

No numerical benchmark results are reported in this model card unless measured and documented separately.

Limitations and Risks

• The model can generate inaccurate calculations or unsupported marketing recommendations.
• ROAS does not necessarily represent profit or incremental advertising impact.
• Results can vary between quantization variants.
• Quantization may reduce output quality, especially at lower bit widths.
• Local inference performance depends on hardware, context length, runtime, and GPU offloading.
• The model does not inherently access live advertising dashboards or current campaign data.
• Thai and English performance may vary by task.
• Users should independently validate important financial and campaign decisions.

License

The fine-tuned model's license and redistribution terms must be verified before publication. The base model is available under its own license, and the dataset may have separate terms.

Base model: https://huggingface.co/Qwen/Qwen3-4B

Fine-tuned model: https://huggingface.co/yothinS/Qwen3-4B-ORPO-Advertisement

Dataset: https://huggingface.co/datasets/yothinS/Advertisement_specialist

Do not assume that the licenses of the base model, fine-tuned model, and dataset are interchangeable. Confirm that the license declared in this repository accurately reflects the rights to redistribute the converted weights.

Acknowledgments

• Qwen Team — for the Qwen3-4B base model.
• Hugging Face — for model hosting and the open-source ecosystem.
• Advertisement Specialist Dataset — for the domain-specific preference data.
• llama.cpp and GGUF ecosystem contributors — for local inference and quantized model support.

Author

Developed and published by yothinS.

Fine-tuned model: https://huggingface.co/yothinS/Qwen3-4B-ORPO-Advertisement

Dataset: https://huggingface.co/datasets/yothinS/Advertisement_specialist