simpra/squad_xh

SQuAD isiXhosa

isiXhosa translation of rajpurkar/squad, built so that the answer is always a verbatim substring of its context — the property extractive QA depends on.

Why this exists

Translating question, answer and context independently breaks extractive QA: the translator renders the same phrase differently in each field, so the answer usually is not present in its own passage. An earlier field-wise translation of SQuAD-train measured 41.1% answer-in-context — meaning ~59% of rows asked the model to copy a span that was not there.

Method — marker-based answer projection

1. Mark the answer span inside the source context: Atop the Main Building is ⟦ a golden statue of the Virgin Mary ⟧.
2. Translate the marked context in a single pass, so the answer is translated in context and stays consistent with the surrounding prose.
3. Read the translated answer back out from between the markers, then strip them.
4. Verify: the answer must be a substring of the context, or the row is dropped.

Marker style ⟦ ⟧ was selected by probing four candidates; it survived translation 5/5 while *** survived 0/5.

	answer-in-context
field-wise translation (earlier attempt)	41.1%
this dataset	100.0%

Fields

field	description
id	original SQuAD id
question	translated question
answer	translated answer — guaranteed substring of context
context	translated passage
category	original SQuAD title
answer_en	the untranslated English answer, for reference

Limitations

• Machine translation. Engine: Google Translate web endpoint, which scored chrF++ 50.93 eng→xho on FLORES-200 devtest in our benchmark (NLLB-200: 48.23). Expect translationese; this is not native isiXhosa.
• Rows whose markers did not survive translation were dropped, so this is a filtered subset rather than a full translation of the source split.
• Derived from SQuAD 1.1: every question is answerable, and answers are short spans from Wikipedia paragraphs.

Licence

Inherits CC BY-SA 4.0 from SQuAD.