Kashif786/Sindhi-Wikipedia-RAG-Benchmark-100

🏛️ Sindhi Wikipedia RAG Benchmark (100-Item Grounded Evaluation Suite)

Source Corpus: Kashif786/Sindhi-Wikipedia-Clean-CPT-2026
Generator / Curator: Google Gemini Argon (Frontier LLM) & Computational Linguistics Research
Evaluator Target: RAG Retrieval Pipelines & Sindhi Models (e.g. Kashif786/gemma-3-1b-sindhi-sft-final / sindhi-gemma-4)
Format: Parquet, JSON, and CSV
Status: Pure Benchmark Dataset (Fresh ground-truth test suite, ready for model evaluation)

---

1. Overview & Dataset Profile

This benchmark consists of 100 curated, high-purity evaluation items (SD-WIKI-RAG-001 through SD-WIKI-RAG-100) grounded in 25 verified articles from the Kashif786/Sindhi-Wikipedia-Clean-CPT-2026 corpus.

It is specifically designed to rigorously test and evaluate:

1. Dense, Lexical (BM25), and Hybrid Retrieval over long-form Sindhi encyclopedic texts.
2. Context-Grounded Instruction Following & Faithfulness in Sindhi generative models.
3. Hallucination Suppression & False-Premise Abstention (guardrail against hallucinating unattested facts).

---

2. Evaluation Tasks & Benchmark Structure

The 100 benchmark items are partitioned across 5 distinct evaluation task types:

Task Type (task_type)	Item Count	What It Evaluates in RAG & Sindhi LLMs
Single-Fact Extractive QA (extractive_qa)	35	Exact retrieval and faithful extraction of dates, names, measurements, and definitions in Sindhi.
Multi-Fact Synthesis (multi_fact_synthesis)	20	Aggregating multi-item lists, tables, and timelines across multiple paragraphs without dropping items.
Conceptual Explanation (explanation_summarization)	20	Fluent Sindhi instruction-following when explaining scientific, technical, or literary concepts.
Disambiguation & Nuance (nuance_reasoning)	15	Preserving competing historical theories or multi-source discrepancies (e.g., UNESCO vs. EFT wall lengths of Ranikot).
Unanswerable / Abstention (unanswerable_abstention)	10	Hallucination guardrail: rejecting false premises when the text explicitly states a fact is unknown/unattested.

---

3. Topical Domain Distribution

Domain Category	Item Count	Covered Wikipedia Articles (Sindhi Titles)
Sindhi Literature, Sufism & Culture	21	شاھ عبداللطيف ڀٽائي, سچل سرمست, سنڌي ادب, سنڌي سڀيتا
Geography & Landmarks	20	ڪراچي, سکر, منڇر ڍنڍ, پاڪستان
History & Political Figures	20	ڪلهوڙا شاهي گهراڻو, هوش محمد شيدي, ذوالفقار علي ڀٽو, پاڪستان
Natural Sciences, Astronomy & Biology	15	طبيعيات, طبعي سائنس, سج, ٻلي, ڪڪڙ
Computer Science & Digital Tech	12	ڪمپيوٽر, ٻٽو عددي نظام, سيمي ڪنڊڪٽر ياداشت, ڪنارا ڪاڍ, عددي عڪس, ڊيٽا, ٻٽ
Archaeology & Heritage	12	موهن جو دڙو, رني ڪوٽ
Total Benchmark Items	100	25 Primary Sindhi Encyclopedic Articles

---

4. Benchmark Schema (11 Columns)

Each row in the dataset contains the following fields:

Column Name	Data Type	Description
benchmark_id	string	Unique benchmark identifier (SD-WIKI-RAG-001 to SD-WIKI-RAG-100).
domain_category	string	One of the 6 primary knowledge domains.
task_type	string	Evaluation task category (extractive_qa, multi_fact_synthesis, etc.).
difficulty	string	Complexity level: Easy (30 items), Medium (47 items), or Hard (23 items).
article_title	string	Source article title in pure Perso-Arabic Sindhi (سکر, رني ڪوٽ, ڪمپيوٽر).
source_url	string	Link to the source article in the Wikipedia CPT corpus.
question_sd	string	Grounded evaluation question in Sindhi.
ground_truth_answer_sd	string	Complete, verified factual reference answer in Sindhi.
context_passage_sd	string	Exact grounded excerpt from the article text.
key_entities_sd	string	Semicolon-separated essential entities for recall scoring.
rag_eval_notes	string	Architectural testing notes (e.g. testing bilingual acronyms, false-premise detection).

---

5. How to Load in Python

from datasets import load_dataset

# Load dataset directly from Hugging Face
dataset = load_dataset("Kashif786/Sindhi-Wikipedia-RAG-Benchmark-100")

# Access train split
benchmark = dataset["train"]
print(f"Total benchmark questions: {len(benchmark)}")

# Inspect first item
item = benchmark[0]
print("ID:", item["benchmark_id"])
print("Domain:", item["domain_category"])
print("Article:", item["article_title"])
print("Question (Sindhi):", item["question_sd"])
print("Ground Truth Answer:", item["ground_truth_answer_sd"])
print("Grounded Context:", item["context_passage_sd"][:150], "...")

---

6. Citation & Attribution

@dataset{sindhi_wiki_rag_benchmark_100,
  author = {Ali, Kashif and Computational Linguistics Research Group},
  title = {Sindhi Wikipedia RAG Benchmark 100: Grounded Evaluation Suite Generated by Gemini Argon},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/Kashif786/Sindhi-Wikipedia-RAG-Benchmark-100}
}