At Sourcetable, our mission is to make data analysis as easy and intuitive as possible — whether you’re crunching numbers, cleaning messy data, or uncovering insights buried deep in your spreadsheets. With built-in AI tools, seamless integrations, and an easy-to-use chatbot for talking to your data, we’ve reimagined what spreadsheets can do. 

Today, we’re excited to announce the addition of TabPFN to Sourcetable, giving you a state-of-the-art AI tool for making predictions directly from your files.

What is TabPFN and Why Does It Matter?

TabPFN (Tabular Prior-data Fitted Network) is an AI model developed by Prior Labs that is designed specifically for working with tabular data, like spreadsheets, so it’s perfect for Sourcetable. It’s built to make accurate predictions quickly and to handle the kinds of real-world datasets you work with every day.

What makes TabPFN stand out is how straightforward it is to use. It delivers predictions in seconds, doesn’t require hours of setup or cleanup, and includes confidence scores with every result so you can assess reliability at a glance. It’s also robust enough to handle messy datasets with missing values and outliers, which is pretty handy and can save hours of cleaning and munging data.

Integrated into Sourcetable, TabPFN offers:

  • Quick Predictions: Get insights in seconds without extra tuning.
  • Handles Real-World Data: Works smoothly with datasets that have missing values or outliers.
  • Accessible to All: No coding required to make the most of its capabilities.
  • Confidence Scores: Understand how reliable predictions are, instead of relying on a single answer.

How to Use TabPFN in Sourcetable

We’ve made it simple for you to harness the power of TabPFN:

  1. Start by uploading your file to Sourcetable. Drag and drop a CSV or XLSX file, or connect directly to a database or one of 100+ app integrations.
  2. Open Sourcetable’s AI chatbot and request the TabPFN model. No setup or configuration is required—just select the model and start working.
  3. Ask your questions. For example: “Which biomarkers indicate disease?”, “Which customers are most likely to churn?” or “What’s the predicted revenue for next quarter?” 

What’s the Catch?

  • TabPFN works best for datasets with fewer than 10,000 rows and 500 columns.
  • This is an early model release:
    • Users should reference TabPFN in the conversation.
    • TabPFN won’t have the same depth of features or autonomous capabilities as other models on Sourcetable.

How does it work?

For those curious about what makes TabPFN tick: it’s a transformer-based model specifically optimized for tabular data. Pre-trained on millions of synthetic datasets, it uses in-context learning to deliver predictions in a single forward pass—no iterative training required.

TabPFN’s architecture includes two-way attention mechanisms, capturing relationships across both rows (samples) and columns (features), and it provides Bayesian-like uncertainty modeling, giving you a probability distribution for every prediction. With GPU acceleration, TabPFN is incredibly fast, handling predictions up to 5,000× faster than traditional models like XGBoost.

Source: https://arxiv.org/abs/2207.01848

Try TabPFN in Sourcetable Today

By integrating TabPFN, Sourcetable continues to redefine what spreadsheets can do. Whether you’re a data scientist, managing a project, exploring trends, or solving complex business problems, you now have a powerful new tool to make smarter, faster decisions.

This is an early release, and we’re going to keep adding more capabilities. In the meantime, get after it! Upload your file, activate TabPFN, and start exploring your data today.




Resources: