Datasets:

islamlab
/

islamic-sciences-training

license: cc-by-nc-sa-4.0
pretty_name: islamlab — Islamic Sciences Training Sets
language:
  - ar
task_categories:
  - text-generation
  - text-retrieval
  - question-answering
size_categories:
  - 1M<n<10M
tags:
  - islam
  - arabic
  - classical-arabic
  - continued-pretraining
  - retrieval
  - citation
  - rag
configs:
  - config_name: cpt
    default: true
    data_files:
      - split: train
        path: cpt/train-*.parquet
  - config_name: retrieval
    data_files:
      - split: train
        path: retrieval/train-*.parquet
  - config_name: attribution
    data_files:
      - split: train
        path: attribution/train-*.parquet
  - config_name: eval_retrieval
    data_files:
      - split: test
        path: eval_retrieval/test-*.parquet
  - config_name: eval_attribution
    data_files:
      - split: test
        path: eval_attribution/test-*.parquet

islamlab — Islamic Sciences Training Sets

Training data derived from islamlab/islamic-sciences: text for domain adaptation, retrieval pairs with hard negatives, and citation questions whose answers are read out of the corpus rather than written by a model.

Nothing here is generated. Questions come from a fixed set of templates and every answer is a field already present in the corpus. That buys a narrow dataset in exchange for one that cannot teach a model a fact the sources do not contain.

Subsets

config	split	rows	what it is
`cpt`	train	780,149	Apparatus-free, deduplicated text with a provenance header — 3.21B characters
`retrieval`	train	906,691	A section heading as the query, a passage from that section as the positive, and hard negatives drawn from elsewhere in the same book
`attribution`	train	130,825	Citation questions with exact answers
`eval_retrieval`	test	26,933	Same shape, held-out works only
`eval_attribution`	test	32,895	Same shape, held-out works only

The held-out split

The holdout is at the level of the work, not the row: 118 of the 4,022 works are assigned to evaluation by a hash of their identifier, and no line of their text appears in cpt, retrieval or attribution. A model trained on the train configs has therefore never read the books it is evaluated on. The assignment is deterministic, so it is the same for everyone.

The tasks in `attribution`

task	question	rows (train)
`author`	من مؤلف هذا النص؟	26,165
`title`	من أي كتاب أُخذ هذا النص؟	26,165
`science`	في أي علم يُصنَّف هذا الكتاب؟	26,165
`century`	في أي قرن هجري توفي المؤلف؟	26,165
`locus`	في أي جزء وصفحة يقع هذا النص؟	26,165

messages is a JSON-encoded chat turn — parse it with json.loads. answer holds the exact expected string. The locus task deliberately includes the honest negative: where the source records no pagination, the answer is لم تُثبَت الصفحة في هذه النسخة rather than an invented page.

Retrieval negatives

Negatives come from the same book as the positive, which is what makes them hard: a chapter of al-Ḥāwī al-Kabīr is separated from another chapter of al-Ḥāwī al-Kabīr by subject alone, not by vocabulary, register or period. negatives is a JSON-encoded list of strings.

import json
from datasets import load_dataset

d = load_dataset("islamlab/islamic-sciences-training", "retrieval", split="train")
row = d[0]
negatives = json.loads(row["negatives"])

Where this sits

Dataset	What it holds
`islamlab/islamic-sciences`	The corpus itself — eight sciences, 4,022 works
`islamlab/islamic-sciences-training`	The training products above
`quranlab/` · `risaleinur/`	Qur'an and hadith; the Risale-i Nur corpus

License

Derived from a corpus licensed CC BY-NC-SA 4.0, and carries the same terms. Credits are in SOURCES.md on the corpus, which is where the works themselves are acknowledged.

Citation

@misc{islamlab_islamic_sciences_training,
  title  = {islamlab — Islamic Sciences Training Sets},
  author = {islamlab contributors},
  year   = {2026},
  howpublished = {Hugging Face Datasets}
}