messykenty/ralomi-ja-enhanced

Ralomi JA Enhanced — Development Preview

This repository is an experimental development preview derived from ayousanz/ralomi.

Current status

• Original Ralomi ONNX model included.
• PyTorch graph conversion validated using onnx2torch.
• Maximum absolute difference against ONNX Runtime: 0.000072479 on one synthetic test input.
• Backward propagation verified for 18 trainable parameter tensors.
• No fine-tuning has been performed yet.
• No accuracy improvement is claimed.

Files

• model.onnx: original compatible inference model
• inference.py: inference entry point
• features.py: audio feature extraction
• inference_config.json: inference configuration
• vocabulary.json: tokenizer vocabulary
• models/ralomi-converted-state.pt: converted PyTorch state_dict
• scripts/convert_and_validate.py: conversion and validation

The PyTorch state_dict is not a standalone model. Reconstruction requires the ONNX graph and compatible onnx2torch.

Limitations

• Model accuracy has not yet been independently benchmarked.
• CoreS3 deployment of the converted PyTorch model is untested.
• Training and export workflows remain experimental.

Attribution

Original model: ayousanz/ralomi

Original license: Apache License 2.0

This development repository is not an official upstream release.