GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
yothinS
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Run this model inference with full control and performance in your environment.
Model Details
Model Provider
yothinS
Model Tree
Qwen/Qwen3-4B
Input Modalities
Output Modalities
Supported Functionality
yothinS/Qwen3-4B-ORPO-Advertisement is a fine-tuned language model designed to assist with advertising strategy, campaign performance analysis, and data-driven marketing decisions.
Built on top of Qwen/Qwen3-4B, this model uses Odds Ratio Preference Optimization (ORPO) to align its responses with preferred advertising-related answers using a specialized preference dataset.
The goal is to develop a more domain-focused assistant that can help marketers interpret advertising metrics, diagnose campaign performance, develop marketing strategies, and make better-informed optimization decisions.
These are the model's intended capabilities, not a guarantee that every response will be accurate or optimal.
| Property | Description |
|---|---|
| Model Name | Qwen3-4B-ORPO-Advertisement |
| Hugging Face Repository | yothinS/Qwen3-4B-ORPO-Advertisement |
| Base Model | Qwen/Qwen3-4B |
| Model Architecture | Qwen3 |
| Approximate Parameters | 4 billion |
| Fine-Tuning Method | Odds Ratio Preference Optimization (ORPO) |
| Primary Domain | Advertisement & Performance Marketing |
The model is derived from Qwen3-4B, a pretrained language model developed by the Qwen team.
Original model: https://huggingface.co/Qwen/Qwen3-4B
Odds Ratio Preference Optimization (ORPO) is a preference-alignment method that combines supervised learning with an odds-ratio-based preference objective.
In preference fine-tuning, the model learns to favor a preferred response (chosen) over a less-preferred response (rejected) for a given prompt.
For this project, ORPO is used to encourage responses that better reflect the quality and reasoning patterns represented in the advertising-specialist preference dataset.
The intended outcome is a model that is more consistent with the preferred responses for advertising-related tasks.
The model was fine-tuned using the following dataset:
Advertisement Specialist Dataset
The dataset provides advertising-related preference examples intended to guide the model toward more useful responses in advertisement strategy and analysis.
Dataset repository: yothinS/Advertisement_specialist
The dataset's published card should be consulted for its exact record count, schema, data-generation methodology, and licensing information.
The following details should be updated to match the actual training run.
| Parameter | Value |
|---|---|
| Base Model | Qwen/Qwen3-4B |
| Fine-Tuning Method | ORPO |
| Training Framework | Unsloth |
| Training Dataset | Advertisement Specialist Dataset |
| Training Records | 877 record |
| Epochs | 2 |
| Learning Rate | 5e-5 |
| Maximum Sequence Length | 1024 |
| LoRA / QLoRA | QLoRA |
This model is intended to support marketing professionals, advertisers, performance marketers, business owners, marketing analysts, and AI developers.
Recommended use cases include:
Advertising Strategy Development
Advertising Performance Analysis
Budget and Campaign Optimization
Prompt:
An advertising campaign generated ฿120,000 in attributed revenue with an advertising spend of ฿30,000. Calculate ROAS and explain what the result means.
Expected task:
Calculate ROAS as revenue divided by advertising spend, yielding 4.0x. Explain that each ฿1 of advertising spend generated ฿4 in attributed revenue, while clarifying that ROAS alone does not establish profitability.
Prompt:
A campaign has a high CTR but a low conversion rate. What should I investigate?
Expected task:
Investigate audience-message alignment, landing-page relevance, offer clarity, page speed, conversion tracking, and potential friction in the conversion journey. Treat these as hypotheses to validate rather than confirmed causes.
Prompt:
I want to improve customer acquisition while controlling CPA. What advertising experiments should I consider?
Expected task:
Recommend structured experiments involving audience segments, creative variations, landing pages, and bidding or budget strategies, with appropriate success metrics and safeguards against misleading comparisons.
These examples illustrate intended use. They are not verified outputs from a model evaluation.
Install the required libraries:
bash
pip install -U transformers accelerate torchLoad the model with Hugging Face Transformers:
python
import torchfrom transformers import AutoTokenizer, AutoModelForCausalLM model_id = "yothinS/Qwen3-4B-ORPO-Advertisement" tokenizer = AutoTokenizer.from_pretrained(model_id) model = AutoModelForCausalLM.from_pretrained( model_id, torch_dtype="auto", device_map="auto",) messages = [ { "role": "user", "content": ( "Analyze this advertising campaign: " "Spend = 10000 THB, Revenue = 40000 THB. " "Calculate ROAS and explain its limitations." ), }] text = tokenizer.apply_chat_template( messages, tokenize=False, add_generation_prompt=True, enable_thinking=False,) inputs = tokenizer(text, return_tensors="pt").to(model.device) with torch.no_grad(): outputs = model.generate( **inputs, max_new_tokens=512, do_sample=False, ) generated_tokens = outputs[0][inputs["input_ids"].shape[1]:] response = tokenizer.decode( generated_tokens, skip_special_tokens=True,) print(response)Note: This example assumes the repository contains the complete model weights and tokenizer files required by Transformers. If the repository contains only a LoRA adapter, load the compatible base model and adapter using PEFT instead. The enable_thinking argument should be used only if supported by the tokenizer's chat template.
A reliable evaluation should test the model on held-out advertising problems that were not included in training.
Suggested evaluation dimensions:
No benchmark scores or comparative performance claims are reported in this model card. Evaluation results should be added after testing.
For important marketing or financial decisions, validate recommendations against real campaign data and controlled experiments.
This model should be used for responsible advertising practices. Users should respect privacy, applicable data protection laws, advertising policies, and consumer rights.
Avoid using model-generated recommendations to create deceptive claims, exploit vulnerable audiences, discriminate unfairly, or violate advertising platform policies.
The base model, Qwen/Qwen3-4B, is published under the Apache-2.0 license. This model card provisionally identifies Apache-2.0 for the fine-tuned model; verify that the training dataset's license, redistribution conditions, and all applicable third-party terms permit this designation before publishing.
Base model license: https://huggingface.co/Qwen/Qwen3-4B
Dataset: https://huggingface.co/datasets/yothinS/Advertisement_specialist
Developed and published by yothinS.
Model: https://huggingface.co/yothinS/Qwen3-4B-ORPO-Advertisement
Dataset: https://huggingface.co/datasets/yothinS/Advertisement_specialist
| Intended Tasks | Text generation, marketing analysis, and strategic decision support |
| Languages | English and Thai are intended for use; actual performance may vary |
| Training Dataset | yothinS/Advertisement_specialist |
| Hardware | T4 16GB VRAM |
Creative and Copywriting
Marketing Analytics and Decision Support
pip install -U transformers accelerate torchimport torchfrom transformers import AutoTokenizer, AutoModelForCausalLM model_id = "yothinS/Qwen3-4B-ORPO-Advertisement" tokenizer = AutoTokenizer.from_pretrained(model_id) model = AutoModelForCausalLM.from_pretrained( model_id, torch_dtype="auto", device_map="auto",) messages = [ { "role": "user", "content": ( "Analyze this advertising campaign: " "Spend = 10000 THB, Revenue = 40000 THB. " "Calculate ROAS and explain its limitations." ), }] text = tokenizer.apply_chat_template( messages, tokenize=False, add_generation_prompt=True, enable_thinking=False,) inputs = tokenizer(text, return_tensors="pt").to(model.device) with torch.no_grad(): outputs = model.generate( **inputs, max_new_tokens=512, do_sample=False, ) generated_tokens = outputs[0][inputs["input_ids"].shape[1]:] response = tokenizer.decode( generated_tokens, skip_special_tokens=True,) print(response)