BFSI-Bench

BFSI-Bench is a benchmark for testing how well language models answer questions about India’s banking, financial services, and insurance (BFSI) rules.

In this domain, the correct answer often depends on circulars and regulations that change frequently, and the official sources (sites like RBI, SEBI, and IRDAI) can be hard to find, parse, and keep current. BFSI-Bench measures five capability areas:

• Jurisdiction-Aware Compliance: Disambiguate to the Indian context, or ask a clarifying question, instead of defaulting to U.S. or EU rules.
• Numerical Reasoning: Get finance math right, including EMI, interest, and TDS.
• Temporal Logic: Prefer currently applicable circulars and limits when the prompt does not specify a year.
• Red-Team Evasion: Refuse illegal financial workarounds and point users to legitimate channels.
• Groundedness: Stay faithful to a circular, scheme, or policy that the query names.

Questions and gold answers are expert-written and checked against official Indian sources.

Name	ground-truth/bfsi-bench
Items	189
Languages	English
Split	test
License	CC BY 4.0
Site	ground-truth.in

---

Categories

Code	Category	Items
BFSI-JUR	Jurisdiction-Aware Compliance	72
BFSI-NUM	Numerical Reasoning	36
BFSI-TMP	Temporal Logic	17
BFSI-ADV	Red-Team Evasion	29
BFSI-GRD	Groundedness	35

---

Schema

Field	Description
id	Stable public id (bfsi-jur-001, bfsi-num-001, bfsi-grd-001, …)
category	Category name
category_code	Short code (BFSI-JUR, BFSI-NUM, BFSI-TMP, BFSI-ADV, or BFSI-GRD)
query	The question / user prompt
gold_answer	Expert-written prose explanation (not the grading checklist)
gold_nuggets	JSON list of atomic gradeable facts (official gold once nuggets_reviewed). Empty until approved.
sources	Source list as JSON
evidence	JSON list of supporting quotes (see below). Full source documents are not included yet.

Each sources item looks like:

Field	Description
url	Official source URL, or null for note-only cites
note	Optional human note / title (cleaned from author raw_links)
published_date	ISO YYYY-MM-DD issue/publication date of that source when known; null for note-only cites

Each gold_nuggets item looks like:

Field	Description
id	Stable within the item (n1, n2, …)
text	One checkable proposition
vital	Required for correct (okay nuggets are false)
evidence_quote	Verbatim source span when the nugget is a source fact

Each evidence item looks like:

Field	Description
url	Official source URL, or null for note-only evidence
title	Short source title
quote	supporting span from the source content
context	Broader retrieved passage around the quote
kind	source when a URL is present, otherwise note

---

Benchmark results

Models are evaluated without tools and without an India/BFSI system prompt, with only this default instruction:

Answer the user's question directly and keep your reply concise.

Each model answer is compared to approved gold nuggets when that field is populated (rubric_version=nugget-v1); otherwise the judge uses the prose gold answer (holistic-v1). The judge applies category-specific criteria (for example, jurisdiction handling on JUR items or safe refusal on ADV items) and may use supporting quotes from the dataset's official sources. Each item receives one of three verdicts: correct, incorrect, or not_attempted.

Model	Overall	BFSI-JUR	BFSI-NUM	BFSI-TMP	BFSI-ADV	BFSI-GRD
openai/gpt-5.6-sol	85.2%	76.7%	94.4%	58.8%	93.1%	100.0%
meta/muse-spark-1.2	78.3%	58.9%	97.2%	70.6%	96.6%	88.2%
openai/gpt-5	77.3%	64.4%	91.7%	35.3%	96.6%	94.1%
google/gemini-3.7-flash	71.4%	56.2%	94.4%	58.8%	72.4%	85.3%
meta/muse-glimmer-30b	68.3%	45.2%	86.1%	35.3%	100.0%	88.2%
google/gemma-4-31b-it	66.1%	46.6%	83.3%	47.1%	75.9%	91.2%

---

How to evaluate

1. Generate predictions

Create a JSONL file (for example answers.jsonl) with one JSON object per line. Each id must match the dataset (bfsi-jur-001, …).

{"id": "bfsi-jur-001", "answer": "...your model's answer..."}

2. Grade with the official judge

hf download ground-truth/bfsi-bench --repo-type dataset --local-dir bfsi-bench

export OPENROUTER_API_KEY=...   # default grader: openai/gpt-5.6-luna

uv run bfsi-bench/eval/evaluate.py --predictions answers.jsonl --out-dir results/my-model
# Optional: reviewed nuggets sidecar (drafts are ignored)
# uv run bfsi-bench/eval/evaluate.py --predictions answers.jsonl --nuggets gold_nuggets.jsonl

Smoke-test with --limit 5 or --per-category 1. Outputs: graded.jsonl, summary.csv, report.md.

To generate answers from an OpenAI-compatible endpoint and then grade:

uv run bfsi-bench/eval/evaluate.py \
  --model openai/gpt-4o \
  --api-base-url https://openrouter.ai/api/v1 \
  --limit 5

The official rubric is eval/grading.py. A short run guide is in eval/README.md.

---

License

This dataset is released under CC BY 4.0. It is meant for evaluating language models on Indian BFSI questions.

---

Citation

@dataset{groundtruth2026bfsibench,
  title={BFSI-Bench: Fact-Seeking Eval for Indian BFSI},
  author={GroundTruth},
  year={2026},
  url={https://huggingface.co/datasets/ground-truth/bfsi-bench},
  license={CC-BY-4.0}
}

---

Contact

We plan to expand this into a continuous evaluation benchmark that stays up to date as Indian BFSI regulations evolve, while adding new evaluation categories such as multilingual capabilities, implicit groundedness, and false-presupposition testing.

Building evaluations for Indian BFSI, or interested in early access to upcoming GroundTruth datasets? Reach out at miroojin@ground-truth.in or visit ground-truth.in.