Dataset · google
Google benchmark for LLM factual accuracy
google/FACTS-grounding-public
860 public English examples to evaluate model response fact-grounding.
What's inside
Public part of the FACTS Grounding 1.0 benchmark from Google DeepMind and Google Research. Contains 860 text examples for English question-answering tasks. The dataset is designed to check how strictly language models rely on provided facts and avoid hallucinations.
What you can do with it
- Testing RAG systems for zero hallucinations and context adherence
- Evaluating factual accuracy of large language models
- Comparing models against public Kaggle leaderboard results
- Tasks
- Question answering
- Languages
- English
- Size
- up to 1K rows
- Downloads this month
- 1.9K
- Likes
- 49
How to load it
from datasets import load_dataset
ds = load_dataset("google/FACTS-grounding-public")