betterflow-indictrans2-indic-en-200m-int8

INT8 ONNX export of IndicTrans2 indic-en distilled 200M, redistributed for on-device use in the Betterflow dictation keyboard.

Provenance

Original model	ai4bharat/indictrans2-indic-en-dist-200M — MIT
ONNX INT8 conversion	hari31416/indictrans2-indic-en-dist-200M-ONNX-int8 — MIT
This repo	byte-identical redistribution of that conversion, pinned for reproducible client downloads

⚠️ This is a third-party conversion, not an official AI4Bharat release. Equivalence against the official PyTorch weights has not been verified.

Files

Seven files are read at runtime: encoder_model.onnx + .data, decoder_model.onnx, decoder_shared.onnx.data, model.SRC, dict.SRC.json, dict.TGT.json. decoder_with_past_model.onnx is included for completeness; the client does not use it — the KV-cache round trip across JNI costs more than the compute it saves at dictation lengths.

Usage notes

Source text must be preprocessed as IndicTrans2 expects: transliterated to Devanagari, Indic digits folded to ASCII, entities masked. The two FLORES tags are single entries in dict.SRC.json and must be prepended as ids, never BPE-encoded. The source sequence needs a trailing </s>.

Measured on device

Motorola Edge 40 (Dimensity 8020), beam 5: ~5.9 s per short sentence, 109 MB resident while running.