Raaj0809/VaaniSetu

---

language: - kok license: cc-by-4.0 pretty_name: Vaani-Setu — Konkani Farm-Advisory Q&A size_categories: - n<1K task_categories: - question-answering - text-generation tags: - konkani - low-resource - agriculture - india - adaptive-data - rag

Vaani-Setu — Konkani Farm-Advisory Q&A

A self-improving dataset of farm-advisory question–answer pairs in Konkani (Devanagari), built for the AI Agents Hackathon 2026 Adaptive Data Track. This dataset is the visible output of an agent-driven flywheel — see the project repo for the agent (the "Long-Tail Hunter") that finds the agent's weakest topics and the Adaptive Data platform that keeps upgrading the corpus on every loop.

Why this dataset exists

Commercial LLMs serve Konkani-speaking farmers poorly: the language has ~2.5M speakers but minimal coverage in major training corpora. Farmers asking about local crops (coconut, cashew, paddy, arecanut), local schemes (PM-Kisan, PMFBY), and local pests (coconut rhinoceros beetle, red palm weevil) get generic or English-leaning answers. This dataset is a small, dense, and adaptively improving corpus targeted at exactly that gap.

Schema

Each row is one JSONL object:

Field	Type	Description
prompt	str	A farmer's question, in Konkani Devanagari
completion	str	An advisory answer, in Konkani Devanagari (1–3 sentences)
enhanced_prompt	str	(adapted rows only) Adaption's reasoning-trace–enriched version of the prompt
enhanced_completion	str	(adapted rows only) Adaption's grounding + reasoning-enriched answer
topic	str	one of sowing, pest, irrigation, scheme, mandi
language	str	Konkani
source	str	synthetic, self_corrected, human_corrected, or adapted
version	int	dataset version this row belongs to

How it was built

1. Seed (synthetic) — A small LLM (Gemini 2.5 Flash) drafted ~100 Q&A pairs across the five topics, prompted explicitly to use Konkani Devanagari and Konkan-coast crop/scheme vocabulary. These rows are tagged source=synthetic.

2. Long-Tail Hunter (probe) — A pinned set of 30 evaluation questions (6 per topic) is run through the advisor agent. A different LLM family (Anthropic Claude Haiku) scores every answer 0–10 on correctness, specificity, safety, and language fidelity. Weak topics are clustered.

3. Corrections — Native Konkani speakers correct the worst answers via the agent's UI. Their corrections are tagged source=human_corrected. Where native speakers are unavailable, the LLM self-critique fills in as source=self_corrected (clearly labeled so consumers can filter).

4. Adapt (Adaption) — The merged corpus is uploaded to Adaptive Data with reasoning_traces=True, deduplication=True, and hallucination_mitigation=True. The output is a grounded, reasoning-trace–enriched JSONL that becomes the next dataset version. These rows are tagged source=adapted.

5. Evaluate — Adaptive Data's evaluation returns score_before → score_after → improvement_percent. The hackathon smoke test on a 10-row English subset moved the score from 4.0 → 8.0 (+100%).

6. Loop — Steps 2–5 repeat. Each cycle, the agent's weakest topics get fresh targeted Q&A, and the corpus gets measurably better.

Statistics

Total rows	~100 (seed) + adapted rows per cycle
Language	Konkani (ISO 639-3 kok)
Script	Devanagari
Topics	5 (sowing / pest / irrigation / scheme / mandi)
Avg Devanagari character share	≥ 0.40 (filter threshold)
License	CC-BY-4.0

Limitations

• Small by design — this is a focused agricultural-advisory dataset, not a general Konkani corpus.
• Synthetic-heavy for cold-start. The source field distinguishes synthetic from human-corrected so you can filter.
• Goan / Konkan-coast bias in crop and scheme references (coconut, cashew, arecanut, PM-Kisan, PMFBY). Konkani spoken elsewhere may want different vocabulary.
• Not for fine-tuning safety-critical applications. Farm advisory should be paired with extension-officer review.

Citation

If you use this dataset, please cite the project and credit Adaptive Data:

@misc{vaani-setu-konkani,
  title     = {Vaani-Setu: A self-improving Konkani farm-advisory dataset},
  author    = {Bagrecha, Mahi and Dhole, Atharv and Choudhari, Raj},
  year      = {2026},
  note      = {Built with Adaptive Data (https://adaptionlabs.ai) for the AI Agents Hackathon 2026}
}

Credit

This dataset is produced by Vaani-Setu and adapted by Adaptive Data — the 500-credit hackathon grant funds the data-flywheel loop that makes this corpus measurably better every cycle.

Team bumblebee · AI Agents Hackathon 2026 · Adaptive Data Track