IndiaFinBench

An Evaluation Benchmark for Large Language Model Performance on Indian Financial Regulatory Text

Rajveer Singh Pall — Gyan Ganga Institute of Technology and Sciences, Jabalpur, India

---

Dataset Summary

IndiaFinBench is, to our knowledge, the first publicly available evaluation benchmark for assessing large language model (LLM) performance on Indian financial regulatory text. Existing financial NLP benchmarks draw exclusively from Western corpora — SEC filings, US earnings reports, and English-language financial news — leaving a significant gap in coverage of non-Western regulatory frameworks.

IndiaFinBench addresses this gap with 406 expert-annotated question-answer pairs drawn from 192 regulatory documents sourced directly from the Securities and Exchange Board of India (SEBI) and the Reserve Bank of India (RBI), spanning documents from 1992 to 2026.

The benchmark covers four task types that probe distinct reasoning capabilities required for Indian regulatory text:

Task Type	Code	Items	Description
Regulatory Interpretation	REG	174	Identify correct rules, thresholds, or scope from regulatory passages
Numerical Reasoning	NUM	92	Perform arithmetic over figures embedded in regulatory text
Contradiction Detection	CON	62	Determine whether two regulatory passages contradict each other
Temporal Reasoning	TMP	78	Order regulatory events, identify which circular was operative at a given time
Total		406

---

Key Results

Twelve models evaluated under zero-shot conditions on the full 406-item benchmark:

Model	REG	NUM	CON	TMP	Overall
Gemini 2.5 Flash	93.1%	84.8%	88.7%	88.5%	89.7%
Qwen3-32B	85.1%	77.2%	90.3%	92.3%	85.5%
LLaMA-3.3-70B	86.2%	75.0%	95.2%	79.5%	83.7%
Llama 4 Scout 17B	86.2%	66.3%	98.4%	84.6%	83.3%
Kimi K2	89.1%	65.2%	91.9%	75.6%	81.5%
LLaMA-3-8B	79.9%	64.1%	93.5%	78.2%	78.1%
GPT-OSS 120B	79.9%	59.8%	95.2%	76.9%	77.1%
GPT-OSS 20B	79.9%	58.7%	95.2%	76.9%	76.8%
Gemini 2.5 Pro	89.7%	48.9%	93.5%	64.1%	76.1%
Mistral-7B	79.9%	66.3%	80.6%	74.4%	75.9%
DeepSeek R1 70B	72.4%	69.6%	96.8%	70.5%	75.1%
Gemma 4 E4B	83.9%	50.0%	72.6%	62.8%	70.4%
Human Baseline (non-specialist)	55.6%	44.4%	83.3%	66.7%	60.0%

All models substantially outperform the non-specialist human baseline. Numerical reasoning is the most discriminative task (35.9 percentage-point spread across models).

---

Dataset Details

Source Documents

Source	Documents	Types
SEBI (sebi.gov.in)	92	Circulars, master circulars, regulations, orders
RBI (rbi.org.in)	100	Circulars, monetary policy statements, master directions
Total	192

Difficulty Distribution

Difficulty	Items	Description
Easy	160 (39.4%)	Single-step extraction from context
Medium	182 (44.8%)	Multi-clause reasoning or calculation
Hard	64 (15.8%)	Multi-instrument tracking or complex arithmetic

Splits

The dataset is split into test (324 items, 79.8%) and dev (82 items, 20.2%).

Split	Items
test	324
dev	82
Total	406

---

Data Fields

Field	Type	Description
id	string	Unique item identifier (e.g., REG_001, NUM_042)
task_type	string	One of: regulatory_interpretation, numerical_reasoning, contradiction_detection, temporal_reasoning
difficulty	string	One of: easy, medium, hard
source	string	Regulatory body: SEBI or RBI
context	string	Regulatory passage(s) provided to the model (80–500 words). For contradiction detection items, contains Passage A and Passage B separated by a delimiter
question	string	The question to be answered from the context
reference_answer	string	Gold-standard reference answer
source_document	string	Filename of the source regulatory document

---

Annotation and Validation

All 406 QA pairs were authored by a domain expert in Indian financial regulation. Every item was individually reviewed to ensure:

• The answer is unambiguously derivable from the provided context
• The question has exactly one correct answer
• The context is sufficient without external knowledge

Model-based secondary validation (LLaMA-3.3-70B, 150-item subset): 90.7% agreement, κ = 0.918 on contradiction detection.

Human inter-annotator agreement (second human annotator, 60-item sample): 76.7% overall agreement, κ = 0.611 for contradiction detection (substantial agreement per Landis & Koch 1977).

---

Evaluation Protocol

Models are evaluated under zero-shot, context-only conditions. The scoring pipeline applies four stages in sequence:

1. Exact match after case-normalisation and punctuation stripping
2. Fuzzy token match using RapidFuzz token_set_ratio ≥ 0.72
3. Numerical extraction match for items where extracted number sets agree
4. Yes/No match for contradiction detection (leading word comparison)

Full evaluation code and all model predictions are available at: https://github.com/rajveerpall/IndiaFinBench

---

Limitations

• All evaluation is zero-shot; few-shot or chain-of-thought prompting may improve performance
• The benchmark does not currently cover Hindi–English code-switched regulatory text
• Coverage is limited to SEBI and RBI; extension to IRDAI, PFRDA, and commodity regulation is planned
• The benchmark evaluates short extractive responses, not longer-form regulatory reasoning or document summarisation
• The dataset is a snapshot of documents as of early 2026; regulatory frameworks evolve continuously

---

Citation

If you use IndiaFinBench in your research, please cite:

@article{pall2025indiafinbench,
  title={IndiaFinBench: An Evaluation Benchmark for Large Language Model Performance on Indian Financial Regulatory Text},
  author={Pall, Rajveer Singh},
  journal={arXiv preprint},
  year={2025},
  url={https://github.com/rajveerpall/IndiaFinBench}
}

---

License

This dataset is released under CC BY 4.0. All source documents are publicly available from sebi.gov.in and rbi.org.in and carry no copyright restrictions on research use.

---

Contact

Rajveer Singh Pall — rajveer.singhpall.cb23@ggits.net
Gyan Ganga Institute of Technology and Sciences, Jabalpur, India