---

SANSKRITI: A Comprehensive Benchmark for Evaluating Language Models' Knowledge of Indian Culture

Link: arxiv.org/abs/2506.15355

Dataset Description

The SANSKRITI benchmark is the largest dataset created to evaluate Language Models' (LMs) comprehension and reasoning capabilities regarding the rich cultural diversity of India.It addresses the critical need for culturally-aware benchmarks, as the global effectiveness of LMs depends on their understanding of local socio-cultural contexts.

Dataset Composition

SANSKRITI comprises 21,853 meticulously curated question-answer pairs.

*Geographical Coverage: It spans all 28 states and 8 union territories of India. *Cultural Attributes: The dataset covers sixteen key attributes of Indian culture, providing a comprehensive representation of India's cultural tapestry: * Rituals and ceremonies * History * Tourism * Cuisine * Dance and music * Costume * Language * Art * Festivals * Religion * Medicine * Transport * Sports * Nightlife * Personalities

---

Data Distribution

The dataset's questions are systematically categorized into four types and show a balanced distribution across these categories:

Question Types

| Question Type | Description |

| :--- | :--- |

| Country Prediction | Determine the country (e.g., India) based on a cultural statement. |

| Association Based | Identify the cultural context or entity most closely associated with a cultural element. |

| General Awareness | Factual questions about general cultural attributes. |

| State Prediction | Identify the specific Indian state referenced in a cultural statement. |

Geographical Distribution

---

Evaluation and Findings

The SANSKRITI benchmark was used to evaluate leading Large Language Models (LLMs), Small Language Models (SLMs), and Indic Language Models (ILMs). The evaluation was performed using a zero-shot, multiple-choice question (MCQ) format, with accuracy as the sole metric.

Key Performance Trends

• Overall Top Performer: GPT-4o demonstrated the best overall performance.
• Open-Source LLMs: LLAMA-3.1-70B-Instruct achieved the highest accuracy (0.86) among open-source LLMs.
• SLMs: Qwen2-1.5B-Instruct emerged as the best-performing SLM, with a score of 0.74, highlighting that some SLMs can possess better domain-specific knowledge than even larger models.

• Question Type Performance:
  • Models achieved the highest accuracy on General Awareness questions.
  • Models showed the lowest accuracy on State Prediction questions.
  • For Association-Based questions, LLAMA-3.1-70B outperformed all other models, including GPT-4o.
• Cultural Attribute Performance: Models performed notably well on Religion, Medicine, and Cultural Common Sense, but generally struggled with Costumes, Cuisines, and Art.
• Geographical Struggles: LMs significantly struggled with questions pertaining to North-Eastern states (e.g., Sikkim, Arunachal Pradesh, Tripura), as well as Bihar and Jharkhand. They performed better for states with globally recognized cities like Delhi and Maharashtra.

---

Dataset Creation and Methodology

Data Sourcing and Organization

The data was sourced from multiple diverse and reliable platforms to ensure authenticity and depth, some of the most used sources are listed below:

1. Wikipedia
2. Ritiriwaz (for customs and rituals)
3. Holidify (for travel, festivals, cuisines, and landmarks)
4. Arts and Culture
5. Times of India

The information was organized in a structured format: ("state name": "attribute": "scrapped data related to the attribute and state").

Annotation Process

A team of 40 annotators — including native and bilingual speakers from various Indian states — was employed to create and validate the questions.

• The annotators were divided into four specialized sub-teams, each focusing on one of the four question types.
• A rigorous cross-validation process was used, where outputs from each sub-team were reviewed by another sub-team to resolve ambiguities and ensure accuracy.
• Cultural sensitivity and ethical guidelines were prioritized, emphasizing the avoidance of stereotypes and encouraging respectful, inclusive framing.
• Annotators were compensated for both question creation and verification.

---

Limitations and Future Work

The authors acknowledge several limitations of the current SANSKRITI dataset:

1.Limited Scope of Cultural Attributes: Only sixteen attributes are covered, not fully representing the entirety of Indian culture. Future work plans to expand attributes to include things like regional folklore and ecological heritage. 2.Lack of State-Specific Multilingual Queries: The current dataset lacks state-specific multilingual questions. This is a key priority for immediate extension. 3.Absence of Visual Question-Answering (VQA) Tasks: The dataset does not include VQA tasks. Future work plans to build a visual question-answering dataset and a multilingual version. 4.Lack of Diversity in Questions: Questions are multiple-choice and fact-based, not explicitly requiring complex reasoning or causal understanding.

📂 Citation

If you use this dataset, please cite:

@inproceedings{maji2025sanskriti,
  title={SANSKRITI: A Comprehensive Benchmark for Evaluating Language Models’ Knowledge of Indian Culture},
  author={Maji, Arijit and Kumar, Raghvendra and Ghosh, Akash and Anushka and Saha, Sriparna},
  booktitle={Findings of the Association for Computational Linguistics: ACL 2025},
  year={2025}
}