Skip to content

TOFU

A benchmark dataset for machine unlearning

Visit official site
Category AI models

What TOFU is

The dataset is described as a benchmark of fictional question answer pairs for evaluating how LLMs unlearn fine-tuned information.

What you can do with it

  • Unlearn a model at several forget set ratios
  • Evaluate unlearning after fine-tuning
  • Test information removal across models

Model notes

  • The dataset uses 200 fictional authors
  • Question answer format suits chat models
  • Llama2 chat and Phi-1.5 are the codebase targets

Pros & cons

✓ What we like

  • Usable with Llama2, Mistral and Qwen
  • The codebase can be adapted to other models

! What to watch out for

  • The authors in the dataset are fictional
  • The codebase targets Llama2 chat and Phi-1.5 first

FAQ

What is in the dataset?

Question answer pairs built from 200 fictional authors.

Which models are named?

Llama2, Mistral, Qwen and Phi-1.5 are named.

Last reviewed: 2026-11-15

More LLM API platform tools

View all →

How we review