ekhodzitsky/rupunct-small-onnx

RUPunct_small — INT8 ONNX

INT8-quantized ONNX export of RUPunct/RUPunct_small (MIT) — Russian punctuation + capitalization restoration as token classification.

Packaged for use in the gigastt speech-to-text server: it runs via ONNX Runtime (no Python at runtime) with the Rust tokenizers crate, restoring punctuation/casing on the plain rnnt head's bare lowercase output, e.g. шестьдесят тысяч тенге сколько будет стоить → Шестьдесят тысяч тенге, сколько будет стоить?.

Files

• rupunct_small_int8.onnx — dynamic-INT8 graph (DynamicQuantizeLinear + MatMulInteger), ~29 MB.
• tokenizer.json, config.json, special_tokens_map.json, tokenizer_config.json — HF WordPiece tokenizer + label map (33 labels).

I/O

• Inputs: input_ids, attention_mask, token_type_ids (int64, [batch, seq]).
• Output: logits [batch, seq, 33]. Take the first sub-token label per word (aggregation first), then apply RUPunct's process_token decode (3 case modes × 11 punctuation classes).

License

MIT, inherited from the upstream RUPunct/RUPunct_small. Export + INT8 quantization only; weights are unchanged in accuracy (fp32 and int8 produce identical labels on test inputs).