Skip to content

Dataset · vtllms

Complex benchmark for models with web search

vtllms/sealqa

Set of evaluation questions (under 1000 pieces) where standard web search yields contradictory or noisy answers.

What's inside

This is a specialized text benchmark of up to 1,000 records, split into subsets of varying difficulty (e.g., seal_0 and seal_hard). It contains factual questions specifically selected so that standard web searches lead to confusing or unhelpful results.

What you can do with it

  • Testing RAG systems and language models with web search access for resilience against information noise
  • Comparing the ability of different models to find truth in contradictory sources
Tasks
Question answering
Size
up to 1K rows
Downloads this month
10.1K
Likes
34

How to load it

from datasets import load_dataset

ds = load_dataset("vtllms/sealqa")