amiralifirouzi/jev-data-persian

Laya-JEV-fa: Persian Typed Judgment Decisions

Laya-JEV-fa is a Persian-translated and sampled version of tasksource/tasksource-jev-typed-decisions, prepared for research, training, and evaluation of Persian language models as part of the LAYA project.

این دیتاست نسخه‌ای فارسی‌شده و نمونه‌گیری‌شده از مجموعه‌ی اصلی Typed Judgment Decisions است که با هدف پشتیبانی از پژوهش، آموزش و ارزیابی مدل‌های زبانی فارسی تهیه شده است.

هر نمونه، بسته به نوع وظیفه، شامل یک زمینه (state)، یک پرسش (question) و در صورت نیاز مجموعه‌ای از گزینه‌ها (options) است. پاسخ هدف با استفاده از ستون target مشخص می‌شود. ستون‌های اصلی انگلیسی برای حفظ قابلیت ردیابی و مقایسه با داده‌های مبدأ نگه داشته شده‌اند.

The dataset preserves the original English fields alongside their Persian counterparts, enabling bilingual analysis, translation evaluation, and comparison between original and translated inputs.

Dataset Overview

Property	Description
Dataset name	Laya-JEV-fa
Primary language	Persian (Farsi)
Original dataset	tasksource/tasksource-jev-typed-decisions
Final dataset size	119,106 rows
Source coverage	662 source datasets
Translation model	tencent/Hy-MT2-7B
Translation framework	vLLM
Main use cases	Text classification, judgment, reasoning, and model evaluation
Project	LAYA
Maintainer	Amirali Firouzi

1. Dataset Structure

این دیتاست شامل ستون‌های اصلی داده‌ی مبدأ و ستون‌های فارسی متناظر است.

Column	Description
state	Original context or state
question	Original question or decision task
options	Original answer options
target	Target vector indicating the correct answer or target label
state_fa	Persian translation of the context
question_fa	Persian translation of the question
options_fa	Persian translations of the answer options
answer_fa	Persian answer derived from the target and corresponding options
mt_untranslated	Flag indicating samples that could not be translated or retained an untranslated input
kind	Task type, such as choice, score, or noul
variant	Sample variant, such as direct or criteria_permutation
group_id	Identifier for related samples
source	Original data source
license	License recorded for the corresponding source
license_use	Source-provided license usage information

Important Notes

• Original English columns are retained for traceability and comparison with the source dataset.
• The order of elements in options_fa must remain aligned with the corresponding options in options.
• For one-hot target vectors, the active position identifies the correct option. For example, [0, 1, 0] identifies the second option.
• answer_fa is derived from the target and the corresponding Persian option where applicable. Structural consistency does not guarantee semantic translation accuracy.
• kind and variant preserve the original task and sample-type identifiers.
• Samples with kind=noul should be interpreted according to the original task definition rather than assumed to be ordinary multiple-choice examples.

Important: Do not reorder options or options_fa without updating the corresponding target representation.

2. Sampling Methodology

نمونه‌گیری با مقدار seed=42 انجام شده است تا فرایند انتخاب نمونه‌ها قابلیت بازتولید داشته باشد.

Metric	Value
Original dataset size	1,028,298 rows
Number of source datasets	662
Final sampled dataset	119,106 rows
Unique groups (group_id)	105,802
Duplicate rows removed	894
Source coverage	662 / 662 (100%)
Maximum observed source share	Approximately 1.7%

The sampling procedure was designed to improve source diversity and reduce over-representation of individual sources.

Target Distribution by Task Type

kind	Target count
choice	60,000
score	30,000
noul	30,000

These are the intended sampling targets, not verified final counts. Their sum is 120,000, while the final dataset contains 119,106 rows. Actual per-category counts should be calculated from the published dataset.

Target Distribution by Variant

variant	Target proportion
direct	60%
criteria_permutation	10%
label_verification	10%
packed_derived	8%
paired_text_format	6%
instruction_paraphrase	4%

These percentages describe the intended sampling distribution and may differ from the final observed distribution.

3. Machine Translation

متون با استفاده از مدل زیر به فارسی ترجمه شده‌اند:

• Model: tencent/Hy-MT2-7B
• Framework: vLLM
• Temperature: 0
• Prompting: Based on the model's documented Model Card instructions

Translation Statistics

Metric	Value
Unique texts processed after deduplication	152,289
Reported processing time	Approximately 5 hours and 40 minutes
Additional translation retries (echo-retry)	7,838
Texts longer than 20,000 characters left untranslated	236

The translation pipeline included deduplication, retry handling, and post-processing to improve consistency.

این ترجمه به‌صورت ماشینی انجام شده است و بازبینی انسانی کامل برای تمام نمونه‌ها تضمین نشده است. بنابراین، ممکن است برخی خروجی‌ها از نظر روانی، دقت معنایی یا حفظ ساختار متن نیازمند بررسی بیشتر باشند.

4. Post-processing

Several post-processing steps were applied to improve structural consistency and preserve important task labels.

Label Consistency

A deterministic label dictionary was used for selected labels to maintain consistency between translated options and their corresponding answers. Examples include mapping True to «درست» and handling known labels from datasets such as PolitiFact.

Recovery of Technical Labels

When machine translation produced malformed mixed-language labels, the original English technical label was retained or restored where applicable.

Unicode Normalization

The following normalizations were applied:

• Arabic ي → Persian ی
• Arabic ك → Persian ک
• Persian digits → Latin digits

Preservation of Technical Identifiers

Some technical labels and identifiers intentionally remain in English, including not_entailment, Bank77 category codes, and SemEval relation labels.

This preserves compatibility with the original dataset and reduces the risk of changing label semantics.

5. Translation Coverage and Quality Statistics

The following figures are reported for the final dataset.

Metric	Count	Percentage
Non-empty question_fa	119,106 / 119,106	100%
Non-empty state_fa	117,651 / 119,106	Approximately 98.8%
Non-empty answer_fa	119,106 / 119,106	100%
Rows with mt_untranslated=True	907	Approximately 0.76%

Note: Non-empty fields do not necessarily indicate semantically correct translations. In particular, answer_fa is derived from the target and corresponding options, so its completeness is not an independent measure of translation quality.

Samples marked with mt_untranslated mainly include procedural text or content based on randomly generated vocabularies.

6. Recommended Use Cases

این دیتاست می‌تواند برای پژوهش‌ها و کاربردهای زیر استفاده شود:

• آموزش و تنظیم دقیق مدل‌های زبانی فارسی
• طبقه‌بندی متن و انتخاب پاسخ
• ارزیابی استدلال و تصمیم‌گیری متنی
• بررسی انتقال دانش از داده‌های انگلیسی به فارسی
• مقایسه‌ی عملکرد مدل‌ها روی متن اصلی و ترجمه‌شده
• توسعه‌ی مجموعه‌های ارزیابی برای مدل‌های زبانی فارسی

The dataset is intended as a research resource. Users should validate its suitability for their particular task, especially when semantic accuracy or high-stakes decisions are involved.

7. Data Splitting and Leakage Prevention

به دلیل وجود نمونه‌های مرتبط در برخی منابع، توصیه می‌شود تقسیم داده به مجموعه‌های آموزش، اعتبارسنجی و آزمون با توجه به group_id انجام شود.

Related samples should ideally remain within the same split to reduce potential data leakage and overly optimistic evaluation results.

Recommended practices:

• Split by group_id rather than randomly splitting individual rows.
• Preserve the alignment between options, options_fa, and target.
• Avoid using evaluation examples during training or prompt development.
• Report whether evaluation was performed on translated text, original text, or both.

8. Limitations

• Machine translation: Translation quality may vary across domains, writing styles, and sentence lengths.
• No guaranteed human review: The dataset should not be treated as a fully human-verified Persian corpus.
• Multilingual source content: Approximately 1.8% of source rows reportedly included languages other than English, including Urdu, Arabic, and Chinese.
• Long texts: A total of 236 texts longer than 20,000 characters were not translated.
• Technical labels: Some identifiers and label codes intentionally remain in English.
• Heterogeneous sources: The original sources may differ in domain, annotation practices, label definitions, and licensing terms.
• Potential data leakage: Random row-level splitting may place related samples in different splits.
• Translation quality versus structural validity: Consistent option ordering and target alignment do not establish that the translated content preserves the original meaning.

Users should independently validate the dataset before relying on it for benchmark claims or downstream applications.

9. Original Dataset and Translation Model

Original dataset:
tasksource/tasksource-jev-typed-decisions

Translation model:
tencent/Hy-MT2-7B

Laya-JEV-fa is a translated and sampled derivative of the original dataset. Please consult the original repository for additional details about the source tasks and their construction.

10. License and Usage

این مجموعه از منابعی با مجوزهای متفاوت تشکیل شده است. اطلاعات مربوط به منبع و مجوز هر ردیف، در صورت وجود، در ستون‌های source، license و license_use نگه داشته شده‌اند.

The original dataset contains heterogeneous licensing information. The presence of a license identifier in a row does not, by itself, establish permission to redistribute that row or the entire derivative dataset.

Before redistributing the dataset, using it commercially, or publishing derivative versions, users should review the applicable terms of the original sources and ensure compliance with any attribution, non-commercial, share-alike, or other conditions.

The metadata field license: other indicates that the dataset does not claim a single standard license for all rows. It should not be interpreted as granting unrestricted use.

📚 Citation

If you use this dataset in research, model training, evaluation, or a public project, please cite this repository:

@misc{firouzi_laya_jev_fa_2026,
  author       = {Amirali Firouzi},
  title        = {Laya-JEV-fa: Persian Typed Judgment Decisions},
  year         = {2026},
  publisher    = {Hugging Face},
  url          = {https://huggingface.co/datasets/amiralifirouzi/jev-data-persian}
}

Please also acknowledge the original dataset and translation model:

• Original dataset: tasksource/tasksource-jev-typed-decisions
• Translation model: tencent/Hy-MT2-7B

Please ensure that citations and licensing requirements for the original sources are respected.

---

👤 Maintainer

Amirali Firouzi

• Hugging Face: @amiralifirouzi
• GitHub: AmiraliFirouzi

Part of the LAYA project.