sbv2-coreml-common / compression_quality.json
AILogDev's picture
Reduce shared BERT size with 8bit blockwise weight storage
eeac9c2 verified
Raw History Blame Contribute Delete
509 Bytes
{
"component": "bert",
"comparison": "original FP32 and int8 block32",
"feature_cases": 24,
"feature_compute_units": "CPU_ONLY",
"maximum_relative_feature_rmse": 0.04466853904118259,
"minimum_token_cosine": 0.9984083980115026,
"listening_check": {
"voice": "JVNV F1 JP-Extra",
"style": "Neutral",
"cases": [
"three sentences",
"normal sentence"
],
"result": "No concerning difference reported",
"scope": "This check does not cover every voice or text."
}
}