Indic-KCC-Agri-Advisory-Benchmark

⚠️ Benchmark only — not agronomic advice. This dataset and its reference answers exist to score language models, not to be used as real farming guidance. KCC references are noisy call-centre transcripts (see Status and caveats); do not act on any answer, reference or candidate, as agricultural advice.

Open-ended agricultural-advisory question answering in 11 Indian languages, built from real farmer questions and the advisory answers given by human agents at India's Kisan Call Centre (KCC). 500 questions were sampled once in English, then translated into the other 10 languages, so every language scores the same 500 underlying questions — cross-language comparisons on this benchmark are apples-to-apples, not confounded by a different question mix per language. Access is gated (see fields at the top of this page); scoring is a reference-based LLM judge, not exact-match.

Why this exists

Existing agricultural-advisory benchmarks are either English-only or use synthetic questions. This one instead asks: can a model give a grounded, safe advisory answer in a farmer's own language, on a real question a farmer actually asked? Building it from real KCC transcripts (rather than synthesizing questions) and holding the question set identical across all 11 languages makes cross-language comparisons meaningful — a model that scores well in Hindi but poorly in Odia is a real signal, not an artifact of a harder question mix.

The dataset is gated for two separate reasons, both real:

• Benchmark integrity. Gold reference answers sitting in the open get scraped into pretraining corpora, and a model that has memorized this benchmark's answers isn't measuring what the benchmark claims to measure. Gating slows that down — it doesn't eliminate it, since anyone who agrees to the terms gets full access, but it removes the benchmark from passive, unauthenticated crawling.
• Farmer privacy. This is real Kisan Call Centre transcripts — actual farmers' questions and the human agent's replies, not synthetic data. A PII audit found and redacted raw phone numbers/emails in a small number of rows, but the underlying transcripts are still real people's queries. Gating adds a deliberate step before redistribution, rather than leaving it fully open by default.

Dataset structure

• 11 Indian languages: Bengali, English, Gujarati, Hindi, Kannada, Malayalam, Marathi, Odia, Punjabi, Tamil, Telugu
• 500 questions, identical across every language (5,500 rows total)
• Real-world source: Kisan Call Centre farmer queries and the human agent's own reply, not synthetic questions
• Reference-based LLM-judge scoring (correctness, naturalness, groundedness, safety) — see Related repositories for the task configs

Each language is its own config (matching this dataset's per-language folder layout), with a single test split.

By language:

Language	Questions	Script forms (native / romanised / code-mixed)
Bengali	500	180 / 162 / 158
English	500	180 / 162 / 158
Gujarati	500	180 / 162 / 158
Hindi	500	180 / 162 / 158
Kannada	500	180 / 162 / 158
Malayalam	500	180 / 162 / 158
Marathi	500	180 / 162 / 158
Odia	500	180 / 162 / 158
Punjabi	500	180 / 162 / 158
Tamil	500	180 / 162 / 158
Telugu	500	180 / 162 / 158

Script-form distribution is identical across every language — it's a property of the question/translation approach, not resampled per language.

By category (query_type) — 19 KCC categories, dominated by pest/disease questions:

query_type	questions
Plant Protection	246
Cultural Practices	76
Nutrient Management	40
Fertilizer Use and Availability	30
Seeds and Planting Material	27
Field Preparation	19
Seeds	16
Weed Management	8
Agriculture Mechanization	8
Bio-Pesticides and Bio-Fertilizers	8
Varieties	5
Soil Testing	4
Vegetative Propagation and Tissue Culture	4
Nursery Management	2
Sowing Time and Weather	2
Water Management	2
Organic Farming	1
Soil Health Card	1
Water Management Micro Irrigation	1

sector splits the same 500 across two broader groups: Horticulture (333), Agriculture (167). crop goes finer still — 235 distinct crops.

Loading example

from datasets import load_dataset

ds = load_dataset("sthanika-ai/Indic-KCC-Agri-Advisory-Benchmark", "Hindi", split="test")
print(ds[0])

Fields

field	description
id	stable row id
idx	source-row index into the original KCC extract
crop, state, district, query_type, season, sector	KCC metadata for the original query
question_en, answer_en	the original English question and reference answer, exactly as transcribed in the source KCC extract
source_answer_used	answer_en, lightly normalised (casing/spacing/punctuation) before translation. This is the text translation actually ran against; it differs from answer_en on ~41% of rows
question, answer	the question/answer translated into this row's language (identical to the _en fields when language == "en")
target_language / language	this row's language code
script_form	native, romanised, or code_mixed
chrfpp	round-trip chrF++ score of the translation against the English source
qc_pass	True if chrfpp >= 50
back_translation	the translation translated back to English, used to compute chrfpp
error	non-empty if translation/QC hit an error for this row
text	== question; the field an lm-evaluation-harness doc_to_text reads

Every row is included, qc_pass failures too — nothing is silently dropped. If you want a stricter subset, filter on qc_pass == True yourself.

Status and caveats

• 0-shot only. Every row is scored, so there is no held-out pool to draw few-shot exemplars from. If you evaluate with lm-evaluation-harness, do not pass --num_fewshot > 0 against these tasks.
• Scoring is judge-based, not exact-match. Open-ended advisory text has no single correct string, so accuracy-style metrics don't apply. Scoring is a reference-based LLM judge (correctness, naturalness, groundedness, safety), run in two stages so the candidate model and the judge model don't need to be loaded together. Task configs and the judge rubric live in the companion GitHub repo — see Related repositories.
• Judge-as-metric. Scores reflect one judge model's opinion, calibrated by nothing but its own prompt. No human-agreement study ships with this release — run one on a sample before treating judge scores as ground truth.
• Reference answers are noisy. KCC references are call-centre transcripts: terse, sometimes redacted ([PHONE]), occasionally incomplete. The judge prompt tells the judge not to penalise a candidate for being more complete than a noisy reference, but this caps how precise correctness can be.
• Machine-translated corpus. Non-English rows are machine translations flagged at chrF++ ≥ 50, not human translations, and failing rows are kept rather than dropped; residual translation error is inside the benchmark. chrfpp and qc_pass are recorded per row so you can audit or filter this.
• qc_pass == False has two different causes. Most such rows genuinely scored below the chrF++ 50 threshold. A small number (40 of 5,500) instead have error: "skipped_short" and a blank chrfpp — the source text was too short to score at all, not necessarily a bad translation.
• English rows are passthrough, not translated, so en is not distributionally comparable to translation quality in the other 10 languages.

Source data and provenance

Source data: India's Kisan Call Centre transcripts, obtained via the Kaggle mirror sridhargutam/kcc-dataset — that mirror self-declares CC0, but that's the re-uploader's own claim on a re-hosted copy, not authoritative over the government source's confirmed terms (see License below); do not treat this dataset as CC0.

500 English questions were sampled once from the source extract, then machine-translated into the other 10 languages with round-trip chrF++ quality control (see Dataset structure and Fields above). The translation and PII-audit pipeline is open source in the companion GitHub repo — see Related repositories.

Citation

If you use this benchmark, please cite both the original KCC data release and this derived dataset.

Original KCC data: Kisan Call Centre transcripts, Government of India (Ministry of Agriculture & Farmers' Welfare) — distributed via the sridhargutam/kcc-dataset Kaggle mirror and data.gov.in.

This derived dataset:

@dataset{indic_kcc_agri_advisory_benchmark,
  title        = {Indic-KCC-Agri-Advisory-Benchmark},
  author       = {sthanika-ai},
  year         = {2026},
  version      = {1.0},
  publisher    = {Hugging Face},
  url          = {https://huggingface.co/datasets/sthanika-ai/Indic-KCC-Agri-Advisory-Benchmark},
  note         = {Derived from Kisan Call Centre (KCC) transcripts, Government of India, licensed GODL-India; see the License section above}
}

License

GODL-India, confirmed at both the platform and resource level. data.gov.in's own Terms of Use / Policies page states in its footer: "The content published on data.gov.in is owned by the respective Ministry/State/Department/Organization and licensed under the Government Open Data License - India." The specific KCC resource page itself was also checked directly (Catalog Info tab): it lists "Released Under: National Data Sharing and Accessibility Policy (NDSAP)" — the original government open-data policy that GODL-India implements platform-wide — and "Contributor: Ministry of Agriculture and Farmers Welfare"; no conflicting license field was found there.

You must comply with GODL-India's attribution requirement when reusing these files:

This work uses data derived from Kisan Call Centre (KCC) transcripts, Government of India (Ministry of Agriculture & Farmers' Welfare), distributed via data.gov.in and licensed under the Government Open Data License – India (GODL-India). Derived and translated by the Indic-KCC-Agri-Advisory-Benchmark project (sthanika-ai).

The companion GitHub repo's code (evaluation harness, audit/build scripts) is licensed separately under MIT — see that repo's LICENSE.

Related repositories

• sthanika-ai/Indic-KCC-Agri-Advisory-Benchmark — the companion GitHub repo: lm-evaluation-harness task configs, the LLM-judge rubric, and the PII-audit/split-build scripts. This HF repo carries the data itself.
• sthanika-ai/Indic-Agri-Benchmark-Model-Configs — model configs for running this and other Indic-Agri benchmarks.