DEV Community· pybankers·· 3 小時前AI 評分41
Tiny GST：由 430 MB 縮至 13 MB，打造印度 GST 發票離線 AI 模型
From 430 MB to 13 MB: Building an Offline AI Model for Indian GST Invoices
AI 導讀
Tiny GST 是一款在裝置上離線擷取印度 GST 發票結構化資料的 AI 模型，經 INT8 ONNX 量化後大小為 13.17 MB。模型在未見過的供應商版面上取得 64.2% entity F1，單價擷取達 84.6% F1，CGST 金額則達 100% F1，並可生成符合 TallyPrime 格式的 XML。
正文
We built Tiny GST, a compact AI model designed to extract structured data from Indian GST invoices directly on-device.
The model:
- Runs fully offline
- Uses only 13.17 MB after INT8 ONNX quantization
- Achieves 64.2% entity F1 on unseen vendor layouts
- Reaches 84.6% F1 for unit-price extraction
- Reaches 80% F1 for HSN/SAC extraction
- Achieves 100% CGST amount F1
- Generates validated TallyPrime-formatted XML
- Includes invoice deduplication using supplier GSTIN, invoice number, and financial year
The goal is simple: reduce cloud costs, protect sensitive invoice data, and make invoice processing work even in warehouses or rural areas with unreliable internet.
The project is currently an early developer release. The XML workflow has been tested against a mock Tally server; live TallyPrime import testing is still pending.
MachineLearning #Android #OpenSource #India
來源：DEV Community · dev.to