Datasets:

InOurWords
/

muslim-womens-interviews

license: cc-by-4.0
language:
  - en
pretty_name: Muslim Women's Interviews (MWI)
size_categories:
  - n<1K
configs:
  - config_name: default
    data_files:
      - split: train
        path: interviews.jsonl
tags:
  - qualitative-research
  - interviews
  - muslim-women
  - hijab
  - representation
  - llm-evaluation
  - social-science

Muslim Women's Interviews (MWI)

De-identified interviews with 25 Muslim women in the United States, Canada and France — 585 question/answer exchanges, about 144,000 words of interview text.

Released with participant consent for the express purpose of shaping how large language models represent Muslim women.

The dataset

Two views of the same 585 question/answer exchanges:

interviews/MWI-001.md … MWI-025.md — one file per interview. Readable on GitHub, importable into NVivo / ATLAS.ti / MAXQDA as a case document.
interviews.jsonl — the same content as 585 structured rows, for loading.

from datasets import load_dataset
ds = load_dataset("InOurWords/muslim-womens-interviews", split="train")
# 585 rows: segment_id, participant_id, country, question_id, question, answer

ds.filter(lambda r: r["question_id"] == "hijab_journey")   # 27 answers, 24 participants
ds.filter(lambda r: r["country"] == "FR")                  # 73 rows, 3 participants

interviews.jsonl is generated from the Markdown by the script in Regenerating the JSONL below, so it is reproducible rather than a second source of truth — every answer matches its .md verbatim.

Each file looks like this:

---
country: "US"
license: "CC-BY-4.0"
---

# Interview MWI-007

## ORDINARY DAY AND WHERE FAITH SHOWS UP

<!-- segment_id: MWI-007_s00 | question_id: ordinary_day -->

An ordinary day — I am a teacher, a middle school teacher. …

Readable on GitHub and the Hugging Face Hub as-is, and importable into NVivo, ATLAS.ti, MAXQDA or Dedoose as a single case document per participant.

Regenerating the JSONL from the Markdown

The Markdown is structured so you do not have to guess. Every answer sits under an ## heading preceded by a comment carrying two stable ids: segment_id (unique across the corpus) and question_id (the same slug wherever that question was asked, so you can group the answers to a given question).

import glob, re, json

PATTERN = r"## (.+?)\n+<!-- segment_id: (\S+) \| question_id: (\S+) -->\n+(.*?)(?=\n## |\Z)"

def parse(path):
    head, body = open(path, encoding="utf-8").read().split("---", 2)[1:3]
    front = dict(re.findall(r'^(\w+):\s*"?([^"\n]*)"?$', head, re.M))
    return [{"segment_id": sid,
             "participant_id": sid.split("_")[0],
             "country": front.get("country"),
             "question_id": qid,
             "question": q.strip(),
             "answer": a.strip()}
            for q, sid, qid, a in re.findall(PATTERN, body, re.S)]

rows = [r for p in sorted(glob.glob("interviews/MWI-*.md")) for r in parse(p)]

with open("interviews.jsonl", "w", encoding="utf-8", newline="\n") as fh:
    for r in rows:
        fh.write(json.dumps(r, ensure_ascii=False) + "\n")

len(rows)   # 585 — byte-identical to the published interviews.jsonl

Nineteen questions recur across most interviews (ordinary_day, hijab_journey, misconceptions, liberation, submission, belonging …); the rest are participant-specific follow-ups.

Redaction

Identifying details are replaced with typed tokens: [CITY 1], [UNIVERSITY 2], [MASJID 1], [INTERVIEWEE]. Token numbering restarts in each interview — [CITY 1] in MWI-004 and in MWI-021 are different cities, so tokens are not keys you can join across participants.

Retained by design: country names, Quebec, France, the US, Quebec legal and identity vocabulary (Bill 21/94/9, CAQ, CEGEP, laïcité, Québécois), ethnicity and nationality descriptors, and public figures cited as commentary.

Composition


Country	US 14 · Canada 8 · France 3

The corpus deliberately contains disagreement. Any use that flattens these women into a single representative voice is a misuse.

Before you use this

Read DATASHEET.md. In short: n=25, purposively sampled partly through Muslim women's organisations and skewed toward educated, community-involved women; France is only 3 participants; the prose was edited into readable narrative, so it is not suitable for conversational-realism or ASR work; and some participants remain identifiable to a motivated reader despite redaction.

Please do not use this corpus to generate synthetic "Muslim woman" personas, or present any participant as representative of Muslim women generally.

License

CC BY 4.0 — share and adapt, including commercially and for model training, with attribution. See CITATION.cff.