Datasets:
dataset_info:
features:
- name: id
dtype: string
- name: context
dtype: string
- name: questions
list:
- name: id
dtype: string
- name: question
dtype: string
- name: answers
struct:
- name: text
list: string
- name: answer_start
list: int32
- name: metadata
dtype: string
splits:
- name: train
num_bytes: 31517284
num_examples: 4548
- name: validation
num_bytes: 5498260
num_examples: 913
- name: test
num_bytes: 4606961
num_examples: 985
download_size: 34172769
dataset_size: 41622505
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
- split: validation
path: data/validation-*
- split: test
path: data/test-*
license: cc-by-4.0
source_datasets:
- X-LANCE/WebSRC_v1.0
tags:
- decision-model
websrc-spans
WebSRC v1.0 span questions over the visible text of each web page (CC BY 4.0).
WebSRC (Chen et al., EMNLP 2021) has crowd-written questions over 6.4k web page segments; span
answers are the annotated DOM element plus a character offset into its text. Each page's HTML is
converted to visible text (VisibleText: decoded entities, collapsed whitespace, one line per
block element, ' | ' between table cells) and every span answer is relocated in the answer
element's range of that text, verified verbatim. Yes/no questions are not spans and are left out
(the tasksource/websrc yes/no config holds them); unlocatable answers are dropped with a
reason, never turned into nulls. One row per page groups its questions. Native website-disjoint
splits: train and dev (validation) from the train+dev archive, test from the test archive
with the answers released in WebSRC_v1.0_test_answers.json.
Original data: X-LANCE/WebSRC_v1.0, original distribution. Repackaged as parquet for tasksource by scripts/repackage_dataset/web_extraction.py.
Conversion provenance: source revisions, hashes, configuration and exclusions.