finndot/finnai-slm-data

FinnAI SLM Training Data

Synthetic training data for fine-tuning on-device models to parse Indian bank SMS into structured JSON. Created for the FinnDot expense tracker.

Key facts

• 100% synthetic — no real user SMS, no real financial data
• Privacy-safe — can be freely shared, no PII
• Apache 2.0 — use for any purpose including commercial
• Multi-language — English, Hindi, Hinglish + seed templates for Tamil, Telugu, Marathi, Bengali
• Multi-task — SMS extraction (60%), finance coaching (25%), general finance Q&A (15%)

Dataset structure

Split	Examples	SMS	Coach	General
train	16,000	9,600	4,000	2,400
val	1,561	1,254	307	0
test	1,592	1,282	160	150

SMS example

{
  "id": "a1b2c3d4e5f6g7h8",
  "task": "sms",
  "bank": "HDFC",
  "template_id": "upi_debit/HDFC/SWIGGY/0",
  "split_key": "HDFC|upi_debit/HDFC/SWIGGY/0",
  "messages": [
    {"role": "system", "content": "Extract transaction details..."},
    {"role": "user", "content": "HDFC Bank: Rs.499.00 debited from A/c XX1234 to SWIGGY via UPI. Avl Bal Rs.15000.50"},
    {"role": "assistant", "content": "{\"amount\": 499.0, \"merchant\": \"SWIGGY\", \"type\": \"EXPENSE\", \"account\": \"1234\", \"balance\": 15000.5, \"category\": \"Food\"}"}
  ],
  "gold": {"amount": 499.0, "merchant": "SWIGGY", "type": "EXPENSE", "account": "1234", "balance": 15000.5, "category": "Food"}
}

Coach example

{
  "id": "b2c3d4e5f6g7h8i9",
  "task": "chat",
  "bank": "COACH",
  "template_id": "coach/leak/led_001",
  "messages": [
    {"role": "system", "content": "You are a personal finance coach. The user's spending this month: Total: ₹18,240..."},
    {"role": "user", "content": "Where am I spending the most?"},
    {"role": "assistant", "content": "You spent ₹7200.00 on Food — that is 39% of ₹18240.00 this month..."}
  ],
  "must_ground": ["18240.00", "7200.00", "65000.00"]
}

Negative example (non-transaction)

{
  "id": "c3d4e5f6g7h8i9j0",
  "task": "sms",
  "bank": "HDFC",
  "template_id": "neg/otp/HDFC",
  "messages": [
    {"role": "system", "content": "Extract transaction details..."},
    {"role": "user", "content": "123456 is your OTP for HDFC Bank NetBanking. Valid for 5 min. Do not share."},
    {"role": "assistant", "content": "{}"}
  ],
  "gold": {}
}

Data sources

Source	Generator script	Description
Parser gold	ml/data/extract_parser_gold.py	JUnit test cases from 35+ bank parsers
Synthetic SMS	ml/data/generate_synthetic_sms.py	Template-based UPI/NEFT/IMPS/CC/salary/ATM
Finance coach	ml/data/generate_finance_chat.py	Grounded Q&A with synthetic ledgers
General instruct	ml/data/generate_general_instruct.py	Short finance Q&A (SIP, EMI, tax)

Note: The published dataset excludes Nova Pro Indic expansions (per AWS service terms). To regenerate Indic data locally, run python -m data.generate_nova_indic with your own Bedrock access.

Split methodology

• Assignment unit: (bank, template_id) — all variants of a template stay in one split
• Hash: SHA-256(seed=42 + split_key)
• Ratios: ~80% train / 10% val / 10% test
• Train mix: 60% SMS, 25% coach, 15% general (with upsampling)
• Chat eval: Frozen 50-item set, never in training

Supported banks

HDFC, SBI, ICICI, Axis, Kotak, PNB, Bank of Baroda, Union Bank, IDFC First, Canara, Federal, Indian Bank, IOB, Central Bank, Karnataka, South Indian, City Union, DBS, HSBC, IDBI, J&K Bank, AMEX, OneCard, Slice, LazyPay, Juspay, Airtel Payments, Jio Payments, IPPB, Jupiter, Utkarsh, Mashreq, and more.

SMS types covered

Type	Examples
UPI debit/credit	Rs.499 debited to SWIGGY via UPI
NEFT/IMPS	Rs.15000 credited via NEFT from SALARY
Credit card	CC XX1234 used for Rs.2999 at AMAZON
Salary	Salary Rs.65000 credited to A/c
ATM withdrawal	Rs.5000 withdrawn from ATM
Balance enquiry	Avl Bal in A/c XX1234: Rs.25000
OTP (negative)	123456 is your OTP → {}
Promo (negative)	Exclusive offer! Get 10% cashback → {}
KYC (negative)	Complete your KYC by visiting... → {}
Failed UPI (negative)	UPI txn failed for Rs.500 → {}

How to use

from datasets import load_dataset

ds = load_dataset("finndot/finnai-slm-data")

# Training
for row in ds["train"]:
    messages = row["messages"]
    # Use with any chat-template trainer (TRL SFTTrainer, axolotl, etc.)

Reproduce from source

git clone https://github.com/AnomaProAI/Finndot
cd Finndot/ml
python -m data.build_splits --repo-root .. --out data/out --train-size 16000
# Verify: sha256sum data/out/*.jsonl  → compare with ml/data/manifest/SHA256SUMS

Integrity

SHA-256 hashes of all split files are in ml/data/manifest/SHA256SUMS in the FinnDot repo. Verify before training.

Citation

@misc{finnai-slm-data-2026,
  title={FinnAI SLM Training Data: Synthetic Indian Bank SMS for On-Device Extraction},
  author={FinnDot Team},
  year={2026},
  url={https://huggingface.co/datasets/finndot/finnai-slm-data}
}

License

Apache 2.0