Mohaddz/syntheta-preview

Syntheta Al-Maktabah Preview

Access, rights, and intended use

This is a public research preview of 300 synthetic Arabic instruction-tuning examples, published at the repository owner's direction. Public access does not grant permission for redistribution, production use, external model training, or downstream relicensing.

The source corpus is `MohamedRashad/Al-Maktabah-Al-Shamilah`, pinned at revision a943e116291d9018acfe87f0afa3f7a89f8ce007. A corpus-wide source license was unresolved when this snapshot was generated. The run used require_license: false as an explicit evaluation exception. This is neither a license determination nor a grant of rights. Prospective users remain responsible for reviewing source rights before copying, redistributing, training on, or otherwise using the material.

Final dataset contents

data/train.jsonl contains exactly 300 Arabic examples. Every row is answerable and contains an instruction, response, subject, task type, educational-value label, verification score, and bounded provenance identifiers. Raw passages and source_context are excluded.

Observed task distribution in the frozen snapshot:

The configured proposal also included unanswerable examples, but none survived the final curation. The observed table, rather than the proposed task weights, is authoritative for this snapshot.

Source sampling and generation

The source was read with Hugging Face streaming; the complete dataset was not downloaded or materialized. Sampling used a deterministic 10,000-row bounded shuffle, base seed 42, and a maximum of 5,000 emitted source rows per classification window. Classification used epochs 0–23, 33, 34, and 1000. Epoch 33 was also read through post-shuffle offsets 5,000, 10,000, 15,000, 20,000, 25,000, and 30,000, in addition to offset 0. An offset causes preceding stream rows to be read and discarded; it is not a random-access seek. The exact epoch, effective seed, and offset tuples are recorded in metadata/run-summary.json.

Generation and automated verification used OpenRouter as the API provider:

• —Generator: deepseek/deepseek-v4-flash
• —Subject classifier: deepseek/deepseek-v4-pro
• —Grounding verifier: deepseek/deepseek-v4-pro
• —Thinking mode: disabled

Automated gates included source-utility screening, standalone-instruction screening, post-generation subject checks, deterministic evidence validation, model grounding verification, safety checks, Arabic-script checks, and exact normalized-pair deduplication. All 300 frozen rows have passing hybrid grounding verification; the minimum score is 0.90, the mean is 0.9995, and the maximum is 1.00. The audit artifact contains 400 evidence spans and 396 supported verification claims.

These automated scores are model-assisted measurements, not a guarantee of scholarly or factual correctness.

Human curation

All 300 rows in the final snapshot underwent manual Arabic content QA. Review covered standalone usefulness, Arabic fluency, subject consistency, answer entailment, evidence completeness, metadata/index trivia, OCR corruption, and hidden-context or unsupported claims. Failed rows were removed, replacements were generated, and the replacement sets were re-audited in sorted batches. The final frozen set then underwent a deterministic content-and-evidence replay against the retained local source chunks; that replay found no row-level content or evidence blockers. Private mutable operator registries are not included in this upload-safe package.

Manual review was used as an acceptance and replacement process; it was not a scored, multi-annotator study. Consequently, metadata/health-report.json still reports no structured human-quality scores.

One final response received a direct one-word human correction. For sample grounded_sft_efacc51e6a80fff647a28a6e, the unsupported word الناظم was removed; the matching verification claim was updated. audit/governed.jsonl records the action, review timestamp, scope, and before/after response SHA-256 hashes under extra.human_curation. This was the only direct manual text edit; other failed rows were replaced rather than rewritten.

Package and governed audit artifact

The upload-safe package is intentionally limited to:

• —README.md
• —checksums.sha256
• —data/train.jsonl
• —metadata/generation-plan.yaml
• —metadata/subjects.yaml
• —metadata/run-summary.json
• —metadata/health-report.json
• —audit/governed.jsonl

audit/governed.jsonl retains structured provenance, verifier outputs, content hashes, bounded evidence quotes, and the single human-correction record. It excludes raw source text and runtime source_context. The private work/ staging area, prompts, attempts, rejection registries, source chunks, credentials, and logs are not part of this package. checksums.sha256 binds the other seven packaged files.

Limitations

• —The examples remain synthetic and may contain factual, interpretive, attribution, classification, or translation errors despite automated and manual review.
• —Classical Arabic and religious material is context-sensitive. This dataset is not a scholarly, legal, medical, or religious authority.
• —Manual review was a single workflow without blinded double annotation, adjudicator agreement statistics, or independently scored quality labels.
• —The accepted task mix differs from the proposal because filtering and curation reject task types unevenly; no unanswerable rows remain.
• —The bounded streaming windows are reproducible for the pinned revision and recorded controls, but they are not a uniform random sample of the complete corpus.
• —Automated privacy, safety, grounding, and duplicate checks reduce risk but cannot eliminate privacy leakage, memorization, harmful errors, or rights concerns.
• —Empty or missing source-license fields must never be interpreted as permission.

Before any broader use, complete source-license review, qualified Arabic-domain review, privacy and memorization review, and a documented rights-and-use decision.

Loading

from datasets import load_dataset

dataset = load_dataset("Mohaddz/syntheta-preview", data_files="data/train.jsonl")

The repository is publicly accessible; access must not be interpreted as a license.

Corrections and takedown

Report corrections or takedown requests through the Hugging Face repository's owner contact or discussion channel. Include the affected sample id, source content hash or source identifier when available, and the reason. The owner should restrict access or remove affected records promptly, document the decision, regenerate the checksums, and publish a corrected revision. Rights-holder requests take precedence over continued public access.