oliverkinch/multi-wiki-qa-high-quality-subset

multi-wiki-qa-high-quality-subset

A quality-filtered subset of the Danish (da) split of alexandrainst/multi-wiki-qa, a Wikipedia-based extractive question-answering dataset.

Configs

Filtering methodology

Starting from the 5,000 samples in the original Danish split:

1. 1.Span validation -- deterministic check that context[answer_start:answer_start+len(answer)] matches the claimed answer text. Samples with misaligned spans are marked incorrect.
2. 2.LLM verification -- each remaining sample is sent to Qwen-235B with a Danish prompt asking it to judge whether the answer is (a) factually correct given the context, (b) actually present in or derivable from the context, and (c) responsive to the question. The model returns a structured {"verdict": ..., "reason": ...} JSON response.
3. 3.Short-answer filter (for da-short only) -- further filters to answers with at most 3 whitespace-delimited words.

Results

Usage

from datasets import load_dataset

# Full quality-filtered subset
ds = load_dataset("oliverkinch/multi-wiki-qa-high-quality-subset", "da")

# Short-answer subset only
ds_short = load_dataset("oliverkinch/multi-wiki-qa-high-quality-subset", "da-short")

Fields

• —id -- Wikipedia article URL
• —title -- article title
• —context -- passage from the article
• —question -- question about the passage
• —answers -- dict with text (list of answer strings) and answer_start (list of character offsets into context)