Sharing user data? Anonymize it locally and stay GDPR-compliant.

Use a task-specific LLM to remove Personally Identifiable Information from text, so you can share data without compromising user privacy.

December 22, 2025

We have a large dataset containing user information and we need to share it with a shareholder, a collegue from a different department, a third-party service provider, or even publish it for research purposes. However, the data contains Personally Identifiable Information (PII) such as names, email addresses, phone numbers, and other sensitive details that a dozen different regulations prevent us from disclosing.

What do we do? A few options come to mind:

1. Going through the dataset, one row at a time, and manually removing Personally Identifiable Information. Possible, but not a great idea: it's tedious, time-consuming and strongly error-prone.
2. Using regular expressions or rule-based scripts to identify and redact PII. This might (or might not) work for simple patterns like email addresses or phone numbers, but it often misses context and more complex identifiers.
3. Sending the data to a third-party service that specializes in data anonymization. This kind of defeats the purpose of anonymization, as we're still sharing sensitive data with an external party.
4. Using an LLM. This really depends on what we mean: if we're thinking of an LLM API, we're back to option 3; if we're thinking of a local LLM, this might actually work... if we have a GPU and a few days (weeks?) to set up and fine-tune the model on the anonymization task.

None of the options really cut it. The local LLM approach seems the most promising: it would allow us to anonymize the entire dataset at once, locally and quickly. However, we don't have a GPU or the time to fine-tune the model.

The task-specific LLM approach

A related tool that allows us to benefit from the power and privacy of a local LLM, without the need for expensive hardware or time-consuming fine-tuning, are Task-Specific LLMs.

Task-Specific LLMs are small language models that have been pre-trained and fine-tuned on specific tasks, such as text classification, token classification, summarization, etc. Because of their small size they don't require a GPU to run, which makes them ideal for local use on a standard personal computer, and because they have already been fine-tuned on the target task, they can be used out-of-the-box without any additional training. Exactly what we need.

Artifex

Artifex is an open-source Python library which makes it easy to use and fine-tune Task-Specific LLMs for various NLP tasks, including text anonymization. Artifex provides a simple API to load pre-trained Task-Specific LLMs and use them to process text data locally.

Text anonymization with Artifex

Anonymizing text with Artifex is straightforward. Let's suppose we have a dataset in a CSV file, user_data.csv. Each row consists of an ID, a message field and a comments field, both of which may contain Personally Identifiable Information. We want to share the dataset with a third-party, but first we need to anonymize the message and comments fields, removing any PII while preserving the rest of the content.

The dataset looks like this:

id,message,comments
1,"Hello, could you please send the report to our colleague Jane Smith?","Done, I CC'd Tom Lewis as well."
2,"Contact me at (123) 456-7890 for further details.","Thanks, I will do that."
3,"Please call Dr. Emily Johnson at 987-654-3210.","The phone number I have is (987) 654-3210, is that correct?"
...
1000,"My address is 123 Main St, Springfield.","Great, I'll see you there."

If you haven't already, install Artifex with

pip install artifex

Using Artifex to load a pre-trained Text Anonymization model and process the dataset is simple:

import csv
from artifex import Artifex

text_anonymizer = Artifex().text_anonymization

with open("user_data.csv", mode="r", newline="", encoding="utf-8") as infile, open("user_data_anonymized.csv", mode="w", newline="", encoding="utf-8") as outfile:
  reader = csv.DictReader(infile)
  fieldnames = reader.fieldnames
  writer = csv.DictWriter(outfile, fieldnames=fieldnames)
  writer.writeheader()
  for row in reader:
      anonymized_message = text_anonymizer(row["message"])[0]
      anonymized_comment = text_anonymizer(row["comments"])[0]
      row["message"] = anonymized_message
      row["comments"] = anonymized_comment
      writer.writerow(row)

All we have done is import Artifex, load the text anonymization model, read the CSV file row by row, anonymize the message and comments fields, and write the anonymized data to a new CSV file.

The resulting anonymized dataset looks like this:

id,message,comments
1,"Hello, could you please send the report to our colleague ...?","Done, I CC'd ... as well."
2,"Contact me at ... for further details.","Thanks, I will do that."
3,"Please call ... at [MASKED]","The phone number I have is [MASKED], is that correct?"
1000,"My address is [MASKED], [MASKED]","Great, I'll see you there."

As you can see, all Personally Identifiable Information has been replaced with the token [MASKED], while the rest of the content remains intact. If the [MASKED] token is not of our liking, we can customize it by passing the mask_token parameter to the text_anonymizer function:

anonymized_message = text_anonymizer(row["message"], mask_token="...")[0]
anonymized_comment = text_anonymizer(row["comments"], mask_token="...")[0]

which results in:

id,message,comments
1,"Hello, could you please send the report to our colleague ...?","Done, I CC'd ... as well."
2,"Contact me at ... for further details.","Thanks, I will do that."
3,"Please call ... at ...","The phone number I have is ..., is that correct?"
1000,"My address is ..., ...","Great, I'll see you there."

The pre-trained model

The text_anonymization method we used to anonymize our file employs a small, pre-trained language model that was created with Artifex itself. If you would like to know more about the model architecture or capapilities, or what the training process looked like, check out the model's Hugging Face page.

Conclusion

Using a Task-Specific LLM to anonymize text locally is the fastest, easiest and most privacy-preserving way to remove Personally Identifiable Information from datasets and stay GDPR-compliant.

Artifex makes it easy to use not only text anonymization models, but also many other Task-Specific LLMs for various NLP tasks, such as sentiment analysis, text classification, named entity recognition, guardrailing, and more. If you want to learn more about Artifex, feel free to explore the Artifex GitHub repository and the Artifex documentation.