A small test model caught a Transformers bug that stayed 2 months

The Transformers team keeps an open source library of AI models. The team proposes Model Integration Contracts. These are small models with recorded outputs, and they operate on each code change. On July 30, 2026, a change let the UMT5 model see future tokens in training. The bug gave no error, and it stayed in the main branch for almost 2 months.

How it works

A contract has 3 parts. A fixture is a small model with the same layer types and the same tokenizer files. A usage path is an example that operates, such as text generation. Baselines are recorded outputs that a person examines. The contract holds if the example on the fixture gives the baselines again.

The numbers

The team replayed commits from 5 months. The small UMT5 fixture showed the bug on all of them until the repair on September 28. The label log probability changed by 1.22, and training loss by about 0.20. On a test branch, the check took 13.4 seconds on a CPU, without setup.

What is not known

The team says fixtures do not replace full tests. The check worked because it used the default attention path. A test of only the eager variant missed the bug. An LLM drafts the contracts, but a person must examine changed outputs.

Insight: AI drafts the test, a person owns the answer

Copilot Fusion lets a second model check the work of a first model. Here, the Transformers team lets an LLM write the test code, but only a person gives approval for the recorded outputs. This reads as work in two parts: models do the first part, and a baseline that a person examined stays the fixed point that AI cannot move. Teams that let models check models must have one such fixed point.