Document Intelligence finetuning not working as desired

Question

Hi, '

We have finetuned the custom model over document intelligence but the result are no any better than pre built model. we have trained the model over 23 documents having mostly pdf and somewhat 1-2 scan pdf only. we have given 4-5 document of 1 type but still the trained model is struggling to parse the output within the same table format.

Answer 1

Hi Navneet Chaudhary

Welcome to Microsoft Q&A Forum, thank you for posting your query here!

It seems default Custom models are not able to capture document information properly. You can benchmark different models like custom neural model against a quality dataset or test with against a group of models. 

Below are points to consider while training for a custom model.

• Insufficient Training Data: Although 23 documents might seem like a good amount, it might still not be enough for the model to generalize well. Consider increasing the number of training documents, especially if they are diverse in terms of layout and content.
• Quality of Training Data: Ensure that the training data is of high quality and accurately labelled. Poor quality or incorrectly labelled data can negatively impact the model's performance.
• Model Complexity: The custom model might not be complex enough to capture the nuances of the documents. Experiment with different model architectures or hyperparameters to see if performance improves.
• Model Composition: Consider using composed models, where you can combine multiple custom models for different tasks (e.g., classification and extraction). This might help improve overall performance.

kindly refer below documentation: https://learn.microsoft.com/en-us/azure/ai-services/document-intelligence/train/custom-neural?view=doc-intel-4.0.0

Thank You.