FLP Wiki

No. We do not train or fine-tune any models on real user search queries. This page explains how we build and improve the models behind CourtListener's semantic search features, and how your queries are and are not used.

How we train our models

The models behind semantic search are fine-tuned on synthetic data. Real user query never appear in this training data, and are never stored in a model's weights.

How we use real queries

If you have query logging turned on, we may use your logged queries in two limited ways to improve our search engine:

  • We use queries to evaluate how our models perform over time, watch for changes in search quality, and decide when retraining is needed.
  • We compare our synthetic training data with real queries to make sure the two are similar semantically. This keeps the data we train on representative of the searches people actually run.

This lets us deliver search results that reflect our users' needs without ever placing real queries into training data or sending them to any external model.

Where training happens

All model training and fine-tuning happens in secure cloud computing environments that we control. We do not use commercial fine-tuning services, and we never share query data with third parties to train their models.

How to opt out

If you'd rather your queries not be logged or used at all, you can turn off query logging in your profile. For details on what we log and how long we keep it, see Are my search queries logged? and the CourtListener Privacy Policy.

Visit My Profile

4 views Last updated 19 hours, 8 minutes ago
Creator: rachel