Data and quality
Fine-Tuning Explained: Training a Model on Your Documents, Complete Guide
Korea Deep LearningPublished
Fine-tuning means further training a model that has already been trained on large-scale data, using data that matches a specific set of documents and business rules, to raise its performance in that area.
Why general-purpose models wobble on business documents
The problem comes from exceptions, not averages.
General-purpose models perform well on typical documents. But the failures that happen in real business operations occur at exceptional documents, not average ones. This is the case when a stamp overlaps a table, a strikethrough cuts across a row, or a fax header is stacked four layers deep.
Measurements show this difference clearly. When we randomly sampled public administrative documents and measured them, a gap appeared in the accuracy of finding item names, but it widened far more in the accuracy of retrieving the values attached to those names. This is the gap between a model trained on a specific document group and one that is not.
The exact definition of fine-tuning: how it differs from extraction criteria adjustment
Three axes of difference between criteria adjustment and fine-tuning
First, what gets changed differs. Adjusting extraction criteria changes the sentences that tell the model what to look for. Fine-tuning changes the model's internal values.
Second, what is required differs. Criteria adjustment needs business knowledge and a few examples. Fine-tuning needs a considerable amount of labeled training data and takes computing resources and time.
Third, how to roll back differs. For criteria adjustment, reverting the sentence is enough. For fine-tuning, you must keep the previous model version to be able to roll back.
The two methods are not substitutes. In terms of order, trying criteria adjustment first and using fine-tuning only for areas that cannot be solved that way is more cost-effective.
Four things to check before deciding on fine-tuning
Four criteria for the decision
First, check whether criteria adjustment solves the problem. Refining item definitions and changing extraction hints solves many issues on their own. Skipping this step causes unnecessary training cost.
Second, check whether you can create training data. You need enough documents with correct answers attached, and creating those answers takes considerable time.
Third, check whether the document group is stable. In areas where the forms change every month, the target will have changed by the time training finishes.
Fourth, check the personal information policy. To use real business documents for training, de-identification must come first, and the training environment must be a controlled one.
How to apply fine-tuning in practice
Collect data centered on difficult cases
Adding many documents that are already handled well does not raise performance. Collecting failed cases and cases that fell out to exception handling is far more effective.
Review history becomes an asset here. If you collect the items that staff corrected repeatedly, along with their originals, that collection is itself the training target.
Fix the evaluation set first
Before training, set aside documents for evaluation. If you measure performance on documents used for training, the results look better than reality.
Keep the evaluation set separate from training data, and compare against the same evaluation set each time you update the model. Only then can you judge whether it improved.
Fine-tuning in a domestic environment
In Korean finance and the public sector, training data in most cases cannot leave the organization. The training environment must be kept internal, or only de-identified data may be taken out.
There are also document conditions. HWP (Hangul word processor) documents, government forms, and faxed receipts are the actual targets, so a model with a record of training on such documents has an advantage from the start. In addition, verification and approval steps are needed whenever the model is updated to keep operations stable.
Adjusting criteria can be the way instead of additional training
For fixed-format enterprise documents, additional training clearly has an effect. However, if the structure repeats training every time new documents come in, the operational burden keeps piling up. Each time a department is added, you must collect data, label it, and validate it again.
So in practice the two approaches are used separately. When the nature of the document group itself differs, for example when handling specialized forms of a particular industry or documents in another language, additional training is appropriate. In contrast, when items increase or a form is revised for documents of the same nature, adjusting extraction criteria and schema is faster and easier to roll back.
The judgment criterion is the nature of the change. Distinguishing whether the model cannot read something from whether the instructions about what to read are insufficient determines which option to choose.
Frequently asked questions
Is it really necessary?
No. Adjusting extraction criteria alone solves many cases. Use fine-tuning in areas that cannot be solved that way.
How much data is needed?
It depends on the document group and the goal. More important than the count is whether enough difficult cases are included.
Do the documents used for training leave the organization?
If the setup trains in an internal environment, they do not. When reviewing adoption, confirm where the training environment is located.
Could performance actually drop?
Yes. If the model is fitted too closely to a specific document group, performance falls on other documents. Keep the evaluation set fixed and compare against it.
How is updating managed?
Keep the previous version, compare it with the same evaluation set, and apply the update after an approval process.
Does additional training need to happen every time the form changes?
Usually not. For documents of the same nature, adjusting extraction criteria and schema is faster and easier to roll back.
When is additional training necessary?
When the nature of the document group itself has changed. If you newly handle specialized forms of a particular industry or documents in another language, it becomes a candidate for review.
If you tell us which document group currently performs poorly, we will first check whether adjusting the extraction criteria can solve it.
Contact us



