🧾 Indonesian Transaction → Structured JSON
Fine-tuning small open LLMs (QLoRA / DoRA) to extract structured fields from
Indonesian bank-SMS and e-wallet messages into {bank, type, amount, counterparty, datetime, balance, ref}. An objective extraction task with a clear metric, trained
locally on a single RTX 3060 12 GB.
Best result: Gemma-4-E2B DoRA 98.8% whole-record exact-match (+43 pp over zero-shot).
Paste an Indonesian bank-SMS or e-wallet notification and extract structured JSON. Powered by Qwen2.5-3B + DoRA adapter (4-bit, ~6.7 GB).
Pre-computed results from scripts/run_all_experiments.py — all three models, all PEFT methods, 400-example held-out test set.
Cost, latency, and accuracy comparison against a hosted frontier model.
Built by Muhammad Fikri Wahidin · QLoRA / DoRA · PEFT · TRL · bitsandbytes · GitHub