SmartKisan-Finance

A frontier LLM fine-tuned for Indian smallholder-farmer financial advisory, covering the 2025-26 / 2026-27 agricultural-finance stack in English, Tamil and Hindi. Built for the Adaption AutoScientist Challenge 2026 (Finance category, Part 1).

Model details

• Base model: mistralai/Mixtral-8x7B-Instruct-v0.1 (46.7B sparse MoE).
• Method: AutoScientist by Adaption — LoRA SFT (rank 64, α 128, dropout 0.05, all-linear target modules, 4 epochs, lr 2e-4 cosine), data + recipe co-optimized.
• Trained model id: adaption_smartkisan_finance (finetune job ac0d033a-…).
• Dataset: solokingM/smartkisan-finance-dataset — a grounded seed expanded with Adaptive Data (dataset d832d855-…).
• Languages: English, Tamil (தமிழ்), Hindi (हिन्दी).

Measured improvement vs baseline

AutoScientist held-out win rate: the adapted model is preferred 66% vs 34% for the base Mixtral-8x7B-Instruct-v0.1 — a +32-point margin (this is the challenge's headline metric).

Adaptive Data quality (dataset): 6.0 → 8.1 (+35% relative), Grade C → B, percentile 7.2 → 17.8.

Metric	Base (Mixtral-8x7B-Instruct-v0.1)	SmartKisan-Finance
AutoScientist win rate	34%	66%

Supplementary fact-accuracy eval

scripts/eval.py scores a 29-question held-out set (disjoint from training — asserted in code) for exact-value fact accuracy, Tamil-script consistency and ROUGE-L. Run it on a GPU (Mixtral-8x7B needs ~25GB+ in 4-bit) and paste the numbers here before submitting:

Metric	Base	SmartKisan-Finance
Fact accuracy	fill %	fill %
Tamil-script consistency	fill %	fill %
ROUGE-L (vs reference facts)	fill	fill

What it knows

• KCC 2026 — ₹5L limit, ₹2L collateral-free, 4% effective rate on prompt repayment (7% base; 3% prompt-repayment incentive; 1.5% MISS subvention to banks), digital e-KCC application.
• PM-KISAN — ₹6,000/yr in 3× ₹2,000 DBT instalments, e-KYC, exclusions, payment-failure fixes.
• PMFBY Kharif 2026 — 2%/1.5%/5% premiums, coverage, new Wild Animal Attack & Paddy Inundation add-ons, claim filing, 12% delay penalty.
• MSP 2025-26 & 2026-27 — per-quintal rates for all Kharif & Rabi crops, procurement agencies (FCI, NAFED, CCI).
• Mandi literacy — AGMARKNET min/max/modal, eNAM.
• Loan comparison — KCC vs moneylender vs MFI vs gold loan, with numbers; JLG for the landless.
• Input financing & scheme stacking — KCC for inputs, PM-KUSUM, warehouse receipt financing, combining schemes.
• General finance literacy — simple vs compound interest, UPI safety, fraud red flags (call 1930), PMJJBY/PMSBY/APY, budgeting.

How to use

This is a LoRA adapter over mistralai/Mixtral-8x7B-Instruct-v0.1. Load the base model and apply the adapter (PEFT), or use the merged weights if you exported them. A GPU is required (the base is 46.7B params).

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
base = "mistralai/Mixtral-8x7B-Instruct-v0.1"
tok = AutoTokenizer.from_pretrained(base)
model = AutoModelForCausalLM.from_pretrained(base, device_map="auto", load_in_4bit=True)
model = PeftModel.from_pretrained(model, "solokingM/SmartKisan-Finance")

Intended use

Advisory support for Indian small/marginal farmers, CSC operators, Krishi Vigyan Kendra counsellors, and agri-fintech developers.

Limitations & safety

• Knowledge cutoff ~June 2026; MSP and scheme values change — re-verify against official portals.
• Not a substitute for a bank officer or government advisory. Informational only.
• Tamil/Hindi coverage is strong for national schemes; some state-specific variations may differ.
• Verify any loan/insurance decision against pmkisan.gov.in, pmfby.gov.in, RBI and NABARD.

Credits

Dataset built and adapted with Adaptive Data, model trained with AutoScientist, by Adaption — AutoScientist Challenge 2026.