desearch/dataset

1---2license: mit3task_categories:4- question-answering5- text-retrieval6language:7- en8tags:9- search10- benchmark11- news12- twitter13pretty_name: Desearch Benchmark Questions14size_categories:15- 100K<n<1M16configs:17- config_name: web18  data_files:19  - split: train20    path: questions/*.jsonl21- config_name: x22  data_files:23  - split: train24    path: x/*.jsonl25---26 27# Desearch Benchmark Questions28 29Fresh, self-contained benchmark questions for evaluating web and X (Twitter) search.30Regenerated daily from recent news and tweets. Each question is answerable from public31sources within a dated window — there are no answer keys or source URLs in the public data,32so systems have to actually search rather than recall.33 34## Subsets35 36| Path | Lane | Built from |37| --- | --- | --- |38| `questions/` | Web / news | Recent news articles (RSS + news sitemaps) |39| `x/` | X / Twitter | High-signal tweets from the past 24h |40 41Files are one per day: `<YYYY-MM-DD>.jsonl` (web) and `x-<YYYY-MM-DD>.jsonl` (X).42 43## Schema44 45Every row is source-free:46 47```json48{49  "id": "…",50  "question": "…",51  "difficulty": "easy | medium | hard",52  "start_date": "YYYY-MM-DDTHH:MM:SSZ",53  "end_date": "YYYY-MM-DDTHH:MM:SSZ"54}55```56 57`start_date`/`end_date` bound the window in which the question is answerable — use them as a58date filter when searching. Gold answers are kept private and are never uploaded.59 60## Loading61 62```python63from datasets import load_dataset64 65web = load_dataset("desearch/dataset", "web", split="train")66x = load_dataset("desearch/dataset", "x", split="train")67```68 69## Updates70 71Regenerated daily by an open-source generator (news twice daily, X once daily), so the set72grows one file per lane per day.73