dataset / README.md
mirianokradze's picture
Update dataset card
cbc6fd5 verified
|
Raw
History Blame Contribute Delete
1.7 kB
metadata
license: mit
task_categories:
  - question-answering
  - text-retrieval
language:
  - en
tags:
  - search
  - benchmark
  - news
  - twitter
pretty_name: Desearch Benchmark Questions
size_categories:
  - 100K<n<1M
configs:
  - config_name: web
    data_files:
      - split: train
        path: questions/*.jsonl
  - config_name: x
    data_files:
      - split: train
        path: x/*.jsonl

Desearch Benchmark Questions

Fresh, self-contained benchmark questions for evaluating web and X (Twitter) search. Regenerated daily from recent news and tweets. Each question is answerable from public sources within a dated window — there are no answer keys or source URLs in the public data, so systems have to actually search rather than recall.

Subsets

Path Lane Built from
questions/ Web / news Recent news articles (RSS + news sitemaps)
x/ X / Twitter High-signal tweets from the past 24h

Files are one per day: <YYYY-MM-DD>.jsonl (web) and x-<YYYY-MM-DD>.jsonl (X).

Schema

Every row is source-free:

{
  "id": "…",
  "question": "…",
  "difficulty": "easy | medium | hard",
  "start_date": "YYYY-MM-DDTHH:MM:SSZ",
  "end_date": "YYYY-MM-DDTHH:MM:SSZ"
}

start_date/end_date bound the window in which the question is answerable — use them as a date filter when searching. Gold answers are kept private and are never uploaded.

Loading

from datasets import load_dataset

web = load_dataset("desearch/dataset", "web", split="train")
x = load_dataset("desearch/dataset", "x", split="train")

Updates

Regenerated daily by an open-source generator (news twice daily, X once daily), so the set grows one file per lane per day.