A LoRA fine-tuned Gemma 4 E2B (google/gemma-4-E2B-it) that reads raw Indian bank SMS and generates structured 13-field JSON transaction records, merged and quantized to GGUF for on-device inference via llama.cpp.

Built for the Build with Gemma: TFUG Prayagraj hackathon, powering the FlashFlow Sentinel Android app — a 100% offline financial SMS tracker.

Files

gemma4_e2b_q4km.gguf — Q4_K_M quantized, ~3.4GB. This is what the FlashFlow Sentinel app bundles (split into <1GB chunks at build time for Flutter asset loading).

Usage

json

{
  "transaction_type": "debit",
  "amount": 449.0,
  "currency": "INR",
  "date": "2026-06-14",
  "time": "19:32",
  "sender_bank": "HDFC Bank",
  "sender_acc": "XX1234",
  "receiver_bank": null,
  "receiver_acc": null,
  "counterparty_name": "SWIGGY",
  "reference_id": "615243987012",
  "balance_after": 18230.55,
  "is_actionable": true
}

Prompt format (Gemma's native turn format):

<start_of_turn>user
You are a financial data extractor. Parse the bank SMS and return a JSON object with these fields: transaction_type, amount, currency, date (ISO 8601), time, sender_bank, sender_acc, receiver_bank, receiver_acc, counterparty_name, reference_id, balance_after, is_actionable. Use null for absent fields.

SMS: <raw SMS text><end_of_turn>
<start_of_turn>model

Training

LoRA rank/alpha/dropout: 8/16/0.05, target modules scoped to model.language_model.layers.*.(self_attn|mlp).*_proj
100 steps, effective batch 8, bf16, on a 2,500-sample synthetic dataset
Full recipe, dataset generator, and training code: https://github.com/4-alok/flashflow-gemma4

Dataset

Synthetic Indian bank SMS, same generation methodology as 4-alokk/flashflow-bank-sms.

4-alokk/flashflow-gemma4-e2b-bank-sms

FlashFlow Gemma 4 E2B — Bank SMS Extraction

Files

Usage

Training

Dataset

4-alokk/
flashflow-gemma4-e2b-bank-sms