tojpaj/kisan-advisory-multilingual-indic

Kisan Advisory (Hindi / Punjabi / English)

Real farmer questions and Farm Tele Advisor answers from India's government Kisan Call Centre helpline, adapted with AutoScientist, expanded into Hindi and Punjabi, and filtered so that every row provably preserves the agrochemical doses in its source note.

Rows (after dose filtering)	6,232
Language split	2,064 en / 2,726 hi / 1,442 pa
Quality grade	E → C (3.0 → 5.8)
Relative improvement	+93.3%
Percentile	13.8
Median answer length	505 chars
Median expansion vs source	6.5×
Dose fidelity	100% by construction

Why the dose filter exists — the main finding

AutoScientist's quality grade does not detect loss of safety-critical specifics. Measured on 500-row samples of this same source:

config	grade	improvement	doses fully preserved	expansion
length: detailed	B 8.7	+190%	61%	63.7×
length: concise	C 5.7	+90%	49%	5.6×
explicit verbatim blueprint	C 6.1	+52.5%	58%	6.9×

The highest-scoring configuration expanded a 44-character note (told him to spray amistar @ 2ml / l of water) into a 10,277-character "Comprehensive Guide" — and still lost ~35% of the source's doses. Optimising the grade actively selected for the more verbose, less grounded output.

Three separate blueprint formulations failed to enforce dose fidelity (49% → 58% → 61%), including one with an explicit worked example and ABSOLUTE REQUIREMENT framing. The conclusion is that the blueprint cannot be relied on for numeric fidelity — it has to be verified externally. dose_filter.py drops any row whose source numbers do not survive into the enhanced completion, so what ships here is 100% dose-faithful by construction.

This matters because a farmer acting on an altered or missing pesticide dose can damage a crop or be harmed. It is a safety property, not a quality preference.

Source

Kisan Call Centre transcripts (Open Government Data Platform India), via the Kaggle mirror daskoushik/farmers-call-query-data-qa. 178,939 raw pairs → 154,215 usable → 2,500 sampled stratified across seven agronomic topics (pest/disease 27%, weather/irrigation 22%, nutrient 16%, livestock 7%, variety/seed 6%, schemes/credit 3%, other 19%).

The source is genuinely messy, which is the point: 98.8% of questions start lowercase, 78.3% are telegraphic fragments (asking about the nursery management of cabbage), with pervasive typos (aplly, recommanded, protienous).

Honest caveat on language

The KCC corpus is English — only 152 of 178,939 raw rows contain Devanagari — because the call-centre operator transcribed each call into English. The Hindi and Punjabi rows here are AutoScientist translations, not natively-authored text.

Stated plainly rather than glossed. The trade is defensible: real farmer intent, translated, is more useful for an advisory model than natively written text about unrelated subjects. The farmers originally spoke Hindi and Punjabi — the English transcript is the lossy layer — so translating back arguably restores the original modality. But it is a translation, and users should know.

Fields

• question / answer — original KCC pair
• enhanced_prompt / enhanced_completion — AutoScientist output
• reasoning_trace — chain-of-thought (from reasoning_traces recipe; maps to column_mapping.reasoning_trace at training time)

Built for the Adaption AutoScientist Challenge, Part 2, Agriculture. Full ablation and methodology: METHODOLOGY.md.