from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
import torch

# Load model and tokenizer
tokenizer = AutoTokenizer.from_pretrained("DarthGlennium/flan-t5-land-survey-extractor")
model = AutoModelForSeq2SeqLM.from_pretrained("DarthGlennium/flan-t5-land-survey-extractor")

# Move to GPU if available
device = "cuda" if torch.cuda.is_available() else "cpu"
model.to(device)

# Your OCR text here
text = "Your land survey document text..."

# Tokenize input
inputs = tokenizer(text, return_tensors="pt", truncation=True, max_length=768)
inputs = {k: v.to(device) for k, v in inputs.items()}

# Generate prediction
with torch.no_grad():
    outputs = model.generate(
        **inputs,
        max_length=512,
        num_beams=5,
        early_stopping=True,
        no_repeat_ngram_size=3,
        length_penalty=1.0
    )

# Decode result
result = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(result)

Training

The model was fine-tuned on land survey documents with the following configuration:

Base model: FLAN-T5 (Large or Base)
Input max length: 768 tokens
Output max length: 512 tokens
Beam search with 5 beams for generation

Output Format

The model outputs JSON-formatted strings with the extracted fields:

json

{
  "Land Surveyor": "John Doe",
  "Surveyed For": "Jane Smith",
  "Certified date": "2024-01-15",
  "Total Area": "1000",
  "Unit of Measurement": "square meters",
  "Address": "123 Main St",
  "Parish": "St. Andrew",
  "LT Num": "LT-12345"
}

Limitations

Works best with English land survey documents
OCR quality significantly affects extraction accuracy
May require post-processing for JSON parsing

Citation

If you use this model, please cite:

@misc{flan-t5-land-survey,
  author = {DarthGlennium},
  title = {FLAN-T5 Land Survey Information Extractor},
  year = {2025},
  publisher = {HuggingFace},
  url = {https://huggingface.co/DarthGlennium/flan-t5-land-survey-extractor}
}