vehicle-title-t5-small

Vehicle Title to Year/Make/Model (T5-small)

A fine-tuned t5-small sequence-to-sequence model that extracts structured vehicle information (year, make, model) from free-form vehicle listing titles.

Model Overview

Input/Output Format

Input:

extract vehicle info: New 2023 Yamaha YZ450F Monster Energy Edition

Output:

Year: 2023 | Make: Yamaha | Model: YZ450F

Intended Use

• Normalizing marketplace or dealership listing titles into structured fields
• Preprocessing step for downstream analytics or search systems expecting (year, make, model) tuples
• Batch processing of vehicle inventory data

This model targets vehicle title parsing; it is not a general-purpose language model.

Training Details

Data

Fields used from source data:

• onlineTitle (free-text input)
• catalogYear, catalogMake, catalogModel (target labels)

Training Configuration

Hardware

Software Environment

• Python 3.14
• PyTorch 2.9.1
• Transformers < 4.57
• scikit-learn

Evaluation Results

Evaluated on 19,793 held-out validation samples from the private corpus.

Field-wise accuracies are computed by parsing the generated string and comparing each field independently against ground truth.

How to Use

from transformers import T5Tokenizer, T5ForConditionalGeneration

MODEL_ID = "umut-celik/vehicle-title-t5-small"

tokenizer = T5Tokenizer.from_pretrained(MODEL_ID, legacy=False)
model = T5ForConditionalGeneration.from_pretrained(MODEL_ID)
model.eval()

def predict_vehicle(title: str) -> str:
    input_text = "extract vehicle info: " + title
    inputs = tokenizer(
        input_text,
        return_tensors="pt",
        truncation=True,
        max_length=256,
    )
    outputs = model.generate(**inputs, max_length=128)
    return tokenizer.decode(outputs[0], skip_special_tokens=True)

# Example
title = "New 2023 Yamaha YZ450F Monster Energy Edition"
print(predict_vehicle(title))
# Output: Year: 2023 | Make: Yamaha | Model: YZ450F

Parsing the Output

def parse_prediction(text: str) -> dict:
    parts = {}
    for segment in text.split("|"):
        segment = segment.strip()
        if ":" in segment:
            key, value = segment.split(":", 1)
            parts[key.strip().lower()] = value.strip()
    return parts

result = parse_prediction("Year: 2023 | Make: Yamaha | Model: YZ450F")
# {'year': '2023', 'make': 'Yamaha', 'model': 'YZ450F'}

Limitations

• Trained on a single private dealership corpus; may overfit to specific brands, naming conventions, and title styles present in that data
• Titles deviating from training patterns (non-English text, very short titles, missing year information) may produce incorrect or incomplete outputs
• Model accuracy for the model field is lower than year and make; complex or uncommon model names are harder to extract
• Not robust against adversarial inputs; treat outputs as probabilistic predictions rather than validated facts

License

This model is released under the CC-BY-NC-4.0 license. Commercial use is not permitted without explicit authorization.

Citation

@misc{celik2024vehicletitle,
  author       = {Celik, Umut},
  title        = {Vehicle Title to Year/Make/Model Extraction with T5-small},
  year         = {2024},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/umut-celik/vehicle-title-t5-small}}
}

Contact

• Author: Umut Celik
• Email: [email protected]
• GitHub: github.com/umutc
• Affiliation: CUNY College of Staten Island