searchqa

**SearchQA** is a question-answering dataset where answers must be found from multiple web search snippets, testing models' ability to aggregate evidence from noisy real-world sources. ## What Is SearchQA? - **Size**: 140,000+ question-answer pairs - **Source**: Jeopardy! questions with Google search snippets - **Challenge**: Extract answers from 50+ noisy search results - **Context**: Real web data, not curated paragraphs ## Why SearchQA Matters Real-world QA involves searching the web, not reading a single clean document. SearchQA tests robustness to noise and evidence aggregation. ``` SearchQA Structure: Question: "What is the capital of Australia?" Search Snippets (noisy, redundant): 1. "...Sydney is the largest city in Australia..." 2. "...Canberra became the capital in 1913..." 3. "...Melbourne was briefly the capital..." 4. "...The Australian Parliament is in Canberra..." ...50+ snippets Model must: 1. Filter irrelevant snippets 2. Aggregate evidence (Canberra appears multiple times) 3. Distinguish "largest" from "capital" → Answer: Canberra ``` **SearchQA Challenges**: | Challenge | Description | |-----------|-------------| | Noise | Many snippets are irrelevant | | Redundancy | Answer repeated differently | | Distractors | Related but wrong entities | | Length | 50+ documents to process |

Go deeper with CFSGPT

Get AI-powered deep-dives, save terms, and run advanced simulations — free account.

Create Free Account