pretty_name: QIMMA-oriented Arabic MCQ Dataset Preview
language:
- ar
license: other
task_categories:
- question-answering
tags:
- arabic
- multiple-choice
- reinforcement-learning
- synthetic
configs:
- config_name: default
data_files:
- split: train
path: training_ready.jsonl
QIMMA-oriented Arabic MCQ dataset v1.0
This frozen release contains 477 training-ready Arabic multiple-choice
questions: 81 from the reviewed v2 pilot and 396 newly
accepted questions generated with openai/gpt-oss-120b.
Use
Use training_ready.jsonl as the canonical merged training pool. Each record contains
the question, four choices, answer, explanation, source attribution, difficulty,
quality adjudication, and release origin. batch_training_ready.jsonl contains only
the newly added 396 questions.
Quality and isolation
- 500 new candidates were generated.
- 396 passed strict final adjudication.
- 0 crossed the conservative QIMMA-overlap thresholds.
- 0 were flagged as duplicates against the prior pool or batch.
- QIMMA prompt text is not included in this release.
- Answer positions were deterministically balanced after all filtering.
The audit scanned 44,441 benchmark rows using normalized
exact matching, fuzzy matching, and multilingual semantic similarity. See
release_manifest.json for thresholds, distributions, model lineage, and costs.
Files
training_ready.jsonl: canonical 477-item merged dataset.batch_training_ready.jsonl: 396 newly accepted items.release_manifest.json: release counts, distributions, audits, models, and cost.review_adjudicated.csv: final decisions for all 500 new candidates.held_out_or_rejected.jsonl: new candidates excluded by quality review.overlap_excluded.jsonl: items excluded by overlap/duplicate audits, if any.contamination_*: QIMMA similarity audit without benchmark prompt text.dataset_duplicate_*: prior-pool and within-batch duplicate audit.pipeline/: reproducible collection, generation, adjudication, and audit scripts.
Reported OpenRouter cost for the new batch was $1.2916. Modal compute is
not included. Source licensing is recorded per item; preserve attribution.