Skip to content

Dataset · facebook

Meta benchmark for testing theory of mind in LLMs

facebook/ExploreToM

Synthetic dataset from Facebook for evaluating and training LLM ability to understand mental states (Theory of Mind).

What's inside

The dataset contains between 10k and 100k English examples in text and tabular formats. Scenarios are programmatically generated using the A* search algorithm over a domain-specific language to create complex narrative structures. The dataset is designed for stress-testing and training language models in contextual and mental state understanding.

What you can do with it

  • Evaluating LLM capability on complex Theory of Mind tasks
  • Fine-tuning reasoning models in multi-step social scenarios
Tasks
Question answering
Languages
English
Size
10K–100K rows
Downloads this month
899
Likes
49

How to load it

from datasets import load_dataset

ds = load_dataset("facebook/ExploreToM")