sthanika-ai/Bharat-Knowledge-Probe-Benchmark

BKP-500 — Bharat Knowledge Probe

Does your model know where it is?

BKP-500 is a benchmark of things every Indian knows and frontier LLMs routinely fumble — lakh/crore arithmetic, Indian digit grouping, state-specific land units (bigha, katha, guntha...), traditional mass units, the Indian fiscal year, agricultural crop seasons, and structural identifiers (PAN, GSTIN, IFSC, PIN codes).

The evaluation harness that runs a model against this dataset and grades the responses lives in a companion repository: github.com/sthanika-ai/Bharat-Knowledge-Probe-Benchmark.

Why this is a benchmark and not a trivia quiz

Most "India knowledge" evals test trivia, which frontier models are already good at and which contaminates fast. BKP-500 tests locale conventions: the arithmetic, unit, and calendar defaults that are unremarkable to a resident and invisible to a model trained mostly on US/EU text. These fail in a specific, measurable, and quietly dangerous way — a lakh/crore slip is a 100× error in a financial number, and it never looks like a refusal.

The methodological spine is the matched control pair. Every quantitative item has a twin that is arithmetically identical but internationally framed:

	Item	Control twin
Prompt	"A scheme outlay is ₹1.2 lakh crore. Express it in USD billions at ₹83/USD."	"A regional budget is €13,200 million. Express it in USD billions at $1.10/€." (matched multi-step arithmetic — a magnitude conversion plus a rate conversion, no Indian units)
Measures	locale competence + arithmetic	arithmetic only

Locale Gap Δ = accuracy(control) − accuracy(India-framed). That single number separates "bad at math" from "doesn't know where it is" — a model can score 95% on the control and 54% on the India-framed twin, and that 41-point gap is the headline, defensible against the obvious reviewer objection ("your questions are just harder").

Dataset structure

The dataset ships as one JSONL file per category under data/items/, plus one JSONL file of matched control twins under data/controls/. Load whichever slice you need via the Hub configs:

from datasets import load_dataset

# Default config: all 552 core items across the 7 categories, one "train" split
items = load_dataset("sthanika-ai/Bharat-Knowledge-Probe-Benchmark", split="train")

# One category only
c1 = load_dataset("sthanika-ai/Bharat-Knowledge-Probe-Benchmark", "c1_numerals", split="train")

# Just the control twins (534 rows, matched to core items via control_id/pairs_with)
controls = load_dataset("sthanika-ai/Bharat-Knowledge-Probe-Benchmark", "controls", split="train")

Core items and control twins are kept as separate configs rather than combined splits of one config, because they are mirror-image records — an item's pairs_with is always null and its control_id points at its twin, while a control's control_id is always null and its pairs_with points back — and a single Arrow schema can't hold both shapes cleanly across splits. Join them on control_id / pairs_with when you need both together (see Dataset structure).

Categories

Config	Category	Items
c1_numerals	Indian numeral system — lakh/crore ↔ million/billion, digit grouping, mixed-notation arithmetic	111
c2_weights_volumes	Traditional mass/volume units (tola, ser, maund...)	72
c3_land_units	State-specific land units (bigha, katha, guntha, kanal-marla...)	104
c4_agricultural_seasons	Crop seasons (kharif/rabi/zaid) and agricultural marketing years	91
c5_fiscal_year	Indian fiscal year and quarter arithmetic	85
c6_government_schemes	Government scheme identity, entitlements, funding splits	74
c7_structural_identifiers	PAN, GSTIN, IFSC, PIN code, vehicle registration, Aadhaar structure	15
controls	Matched control twins (internationally framed, arithmetically identical)	534
default	All 552 core items concatenated (data/items/all_items.jsonl) — the convenience config load_dataset(...) gives you with no config name	552

1,086 rows total (552 core items across the 7 categories + 534 control twins).

Fields

Every row (core item or control) shares one schema — see data/schema.json for the full JSON Schema definition. The key fields:

Field	Type	Description
id	string	Unique id, BKP-C<category>-<seq>, e.g. BKP-C3-0142. Control ids use the 5xxx range within their category, e.g. BKP-C1-5001.
category	string	One of the 7 categories above, or control_twin subcategory when is_control is true.
subcategory	string	Finer-grained tag within the category (e.g. magnitude_conversion, bigha_family).
prompt	string	The question text, ≤40 words.
state	string | null	2-letter Indian state/UT code, when the item is state-specific.
answer_type	string	numeric, numeric_with_unit, date, date_range, enum, string_normalized, month_set, or clarification.
gold	object	The gold answer; shape depends on answer_type.
tolerance	object | null	Grading tolerance (exact, relative, or absolute) for numeric answers.
accepted_aliases	list[string]	Accepted alternate phrasings for string-normalized answers.
grader	string	Which grader in the eval harness scores this item.
is_control	bool	true for control-twin rows (in data/controls/).
control_id / pairs_with	string | null	Cross-reference linking a core item to its control twin and vice versa.
provenance	list[object]	Source, URL, quote, and confidence for the facts behind the item — see Considerations.
canary	string	Per-item canary string (BKP500-CANARY-<hex>) for contamination detection — see below.
difficulty	int	1–5 author-assigned difficulty tier.
split	string	draft, dev, or test — see Status & caveats.

Status & caveats

This is an active work in progress, not a finished v1 release.

• No public/gated split yet. Every row currently ships in the draft split with its gold answer visible in plain JSONL. The eventual design (100-item public dev set / 400-item gated test set) exists specifically to resist contamination — a model trained on this repo could later "solve" items by memorization rather than locale competence. Treat scores computed against the current corpus as provisional until that split lands.
• Every row carries a unique canary string (BKP500-CANARY-<hex>). If you are building a pretraining corpus and want to honor benchmark-exclusion norms, filter out any document containing one of these strings.
• Some rows — state land units especially — have provenance gaps where an authoritative source wasn't confirmed at authoring time; check provenance[].confidence before treating a fact as settled.

Considerations for using the data

• This is a locale-knowledge probe, not a general trivia set. Item design deliberately favors facts a resident would consider unremarkable (unit conversions, calendar conventions, identifier formats) over facts that require specialized domain knowledge. state/district_scope mark items whose gold answer depends on a specific jurisdiction.
• Volatile items (volatile: true, mostly in c6_government_schemes) have gold answers tied to a specific policy snapshot (as_of date) and can go stale after a Budget or scheme revision.
• Gold answers are visible. Nothing here is currently held out — don't cite scores against this corpus as evidence of a model's uncontaminated locale competence without checking whether the model's training cutoff postdates this dataset's publication.

Source data and provenance

Numeric/unit facts underlying the items are drawn from a hand-curated registry (government notifications, RBI/NSO publications, and state revenue department references) with per-row provenance recorded in each item's provenance field. See the companion GitHub repository for the registry source files and sourcing methodology notes.

Citation

@misc{bkp500,
  title  = {Bharat Knowledge Probe (BKP-500): Does Your Model Know Where It Is?},
  author = {{Sthanika AI}},
  year   = {2026},
  url    = {https://huggingface.co/datasets/sthanika-ai/Bharat-Knowledge-Probe-Benchmark}
}

License

Licensed under CC-BY-NC-4.0. If you reuse or redistribute this data, please attribute it as:

Bharat Knowledge Probe (BKP-500) dataset, © 2026 Sthanika AI, licensed under CC-BY-NC-4.0.

See LICENSE-DATA.md for the full license note. The evaluation harness code (separate repository) is licensed under Apache-2.0.