Model: yothinS/Llama-3.2-3B-ORPO-Advertisement Source: Original Platform
language, license, library_name, base_model, datasets, pipeline_tag, tags
| language | license | library_name | base_model | datasets | pipeline_tag | tags | ||||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
|
llama3.2 | transformers | unsloth/Llama-3.2-3B-Instruct |
|
text-generation |
|
Advertisement Specialist — ORPO Fine-Tuned Model
Model Description
Advertisement Specialist is a domain-specialized language model fine-tuned to assist with advertising strategy, campaign analysis, and performance marketing tasks.
Built on top of unsloth/Llama-3.2-3B-Instruct, this model uses Odds Ratio Preference Optimization (ORPO) with parameter-efficient fine-tuning to encourage more useful, relevant, and structured responses for advertising-related tasks.
The model is designed to support marketers, advertisers, marketing analysts, and AI developers working with digital advertising campaigns.
Intended Applications
- Advertisement strategy development
- Performance marketing analysis
- Advertising campaign diagnosis
- Campaign optimization recommendations
- Audience targeting and segmentation
- Advertising creative and copy analysis
- Conversion rate optimization
- A/B testing interpretation
- Retargeting and customer acquisition strategy
- Marketing KPI interpretation and recommendations
Model Details
| Property | Details |
|---|---|
| Model name | Advertisement Specialist |
| Base model | unsloth/Llama-3.2-3B-Instruct |
| Model architecture | Llama 3.2 |
| Model size | Approximately 3B parameters |
| Fine-tuning method | ORPO |
| Parameter-efficient method | LoRA / QLoRA |
| Training quantization | 4-bit |
| Training environment | Google Colab |
| Target hardware | NVIDIA T4 GPU |
| Training dataset | yothinS/Advertisement_specialist |
| Training objective | Preference optimization for advertising-domain responses |
| Primary languages | English and Thai |
Training Methodology
1. Dataset
The model was fine-tuned using the Advertisement Specialist Preference Dataset.
The dataset uses the ORPO preference format:
- Prompt: An advertising-related question or task.
- Chosen: The preferred response.
- Rejected: A less-preferred response.
This format allows the model to learn preferences between alternative responses rather than relying solely on conventional next-token supervised fine-tuning.
2. Fine-Tuning Approach
The training pipeline combines ORPO with QLoRA to reduce the number of trainable parameters and lower memory requirements.
Key components include:
- Four-bit model loading during fine-tuning.
- LoRA adapters for parameter-efficient updates.
- ORPO preference optimization.
- Gradient accumulation to accommodate limited GPU memory.
- Evaluation on a held-out dataset, when configured.
3. Training Configuration
| Parameter | Configuration |
|---|---|
LoRA rank (r) |
16 |
| LoRA alpha | 16 |
| LoRA dropout | 0.0 |
| Learning rate | 5e-6 |
| ORPO beta | 0.1 |
| Per-device batch size | 1 |
| Gradient accumulation steps | 4 |
| Maximum sequence length | 2048 tokens |
| Optimizer | adamw_8bit |
| Learning rate scheduler | Linear |
| Weight decay | 0.01 |
| Gradient checkpointing | Enabled |
| Training epochs | 2 |
The configuration above documents the intended training setup. Actual run settings should be verified against the saved training configuration and logs.
Intended Use
This model is intended to serve as an AI assistant for advertising professionals, digital marketers, marketing students, and developers building marketing-focused AI applications.
Example tasks include:
- Explaining advertising KPIs such as CTR, CPC, CVR, CPA, and ROAS.
- Diagnosing campaigns with high click-through rates but low conversion rates.
- Comparing advertising strategies across funnel stages.
- Identifying possible causes of campaign underperformance.
- Developing hypotheses for A/B testing.
- Recommending optimization experiments based on supplied campaign data.
Limitations
- Domain specialization: Performance may vary on tasks unrelated to advertising and marketing.
- Numerical accuracy: KPI calculations and financial recommendations should be independently verified.
- Incomplete information: The model may produce uncertain or incorrect conclusions when campaign data is insufficient.
- Causal inference: Correlations in campaign data do not necessarily establish causation.
- Attribution: ROAS and conversion results depend on attribution windows, tracking quality, and channel overlap.
- Preference-data limitations: Response quality is influenced by the coverage, diversity, and quality of the chosen and rejected examples.
- General capabilities: Fine-tuning may affect the base model's general instruction-following abilities. This should be tested rather than assumed to be preserved.
This model should support human decision-making, not replace careful campaign measurement or professional judgment.
Evaluation
The model should be evaluated using both held-out preference data and practical advertising scenarios.
Recommended evaluation dimensions:
| Metric or test | Purpose |
|---|---|
| Evaluation loss | Monitor performance on held-out preference examples |
| ROAS and CPA calculation accuracy | Verify numerical reasoning |
| Campaign diagnosis quality | Assess analytical usefulness |
| Recommendation relevance | Assess whether recommendations address the stated problem |
| A/B testing reasoning | Evaluate experimental design and interpretation |
| General instruction-following | Check for degradation outside the target domain |
Actual evaluation results should be added after reviewing the training logs and testing outputs. No benchmark scores are claimed in this model card.
Usage
Load the merged model using Transformers:
from transformers import AutoTokenizer, AutoModelForCausalLM
import torch
MODEL_ID = "yothinS/Llama-3.2-3B-ORPO-Advertisement"
tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
model = AutoModelForCausalLM.from_pretrained(
MODEL_ID,
torch_dtype="auto",
device_map="auto",
)
messages = [
{
"role": "user",
"content": (
"A campaign spent 10,000 THB and generated "
"30,000 THB in attributed revenue. "
"Calculate ROAS and explain the result."
),
}
]
inputs = tokenizer.apply_chat_template(
messages,
tokenize=True,
add_generation_prompt=True,
return_tensors="pt",
).to(model.device)
with torch.inference_mode():
outputs = model.generate(
inputs,
max_new_tokens=256,
do_sample=False,
)
answer = tokenizer.decode(
outputs[0][inputs.shape[-1]:],
skip_special_tokens=True,
)
print(answer)
Replace YOUR_USERNAME/YOUR_MERGED_MODEL_REPO with the actual Hugging Face repository containing the merged model. For the separate LoRA adapter repository, load the adapter using PEFT and the original base model instead.
Ethical Considerations
Advertising recommendations can affect consumer behavior, budgets, and business decisions. Users should verify factual claims, comply with applicable advertising and privacy regulations, and avoid misleading claims, discriminatory targeting, or manipulative practices.
Acknowledgements
- Base model: Unsloth Llama 3.2 3B Instruct
- Training dataset: Advertisement Specialist Preference Dataset
- Fine-tuning approach: ORPO with LoRA / QLoRA
Disclaimer
This is an independently fine-tuned model. It is not an official Meta or Unsloth release and should not be interpreted as endorsed by either organization.