lastabhi/amar-bge-m3

Amar BGE-M3 (OCR-tuned, ONNX) for on-device search

The meaning-search model of the Amar Vault Android app: it turns a sentence or a page of text read from a photo or document into a 1024-number vector, so a typed question can find pages that say the same thing in different words, in English and Hindi. Built on BAAI/bge-m3.

Folder: bge-m3-ocr-int4/ (568 MB model + 17 MB tokenizer).

File	Size	What it is
bge-m3-ocr-int4.onnx	568.3 MB	the int8-quantised ONNX model; inputs input_ids, attention_mask; output last_hidden_state; opset 17
tokenizer.json	17.1 MB	the tokenizer the app uses
config.json, tokenizer_config.json, special_tokens_map.json	tiny	upstream configuration
model_metadata.json	tiny	as written by the export: 1024 dimensions, max length 128 tokens, languages english and hindi
files.json		every file's name, size and SHA-256, for checking a download

The model is quantised to int8. The file name keeps "int4" from an earlier name and is unchanged so that existing downloads and the app keep working.

Download only what you need

hf download lastabhi/amar-bge-m3 --include "bge-m3-ocr-int4/*" --local-dir amar-bge-m3

Plain HTTPS: https://huggingface.co/lastabhi/amar-bge-m3/resolve/main/bge-m3-ocr-int4/files.json, then each file it lists. Check size and SHA-256 before use. Range requests are supported, so a download can resume.

Licence

BAAI/bge-m3 is released under the MIT licence; this derived ONNX export is released under the same licence. Copyright of the original model remains with its authors (BAAI). No warranty.