PII-Tracer-webgpu / catalog.json
hybridherbst's picture
Point convertedWith at the tool's path in the medikit repository
96a2a8d
Raw History Blame Contribute Delete
2.62 kB
{
"baseModel": "perplexity-ai/PII-Tracer",
"baseRevision": "d25c16f2e57e321f6d2527715c01df9112f956f5",
"convertedWith": "medikit packages/pii-portal/tools/pii-tracer",
"opset": 20,
"inputs": {
"input_ids": "int64 [batch, sequence]"
},
"outputs": {
"logits": "float32 [batch, sequence, 37]",
"sensitivity": "float32 [batch, 1]"
},
"maxSequenceLength": 4096,
"piiTypes": [
"private_person",
"private_email",
"private_phone",
"private_address",
"private_url",
"private_date",
"account_number",
"secret",
"other_pii"
],
"labels": [
"O",
"B-private_person",
"I-private_person",
"E-private_person",
"S-private_person",
"B-private_email",
"I-private_email",
"E-private_email",
"S-private_email",
"B-private_phone",
"I-private_phone",
"E-private_phone",
"S-private_phone",
"B-private_address",
"I-private_address",
"E-private_address",
"S-private_address",
"B-private_url",
"I-private_url",
"E-private_url",
"S-private_url",
"B-private_date",
"I-private_date",
"E-private_date",
"S-private_date",
"B-account_number",
"I-account_number",
"E-account_number",
"S-account_number",
"B-secret",
"I-secret",
"E-secret",
"S-secret",
"B-other_pii",
"I-other_pii",
"E-other_pii",
"S-other_pii"
],
"variants": [
{
"id": "q8",
"default": true,
"file": "onnx/model.onnx",
"externalData": [
"onnx/model.onnx_data"
],
"bytes": 615736663,
"weights": "8-bit MatMulNBits, block 128, integer zero point; int8 embedding; fp32 heads",
"executionProviders": [
"webgpu",
"wasm"
],
"argmaxAgreementVsFp32": 1.0,
"note": "The variant to use in a browser. onnxruntime-web's WebGPU kernels run this one on the GPU."
},
{
"id": "q4",
"default": false,
"file": "onnx/model_q4.onnx",
"externalData": [
"onnx/model_q4.onnx_data"
],
"bytes": 488434478,
"weights": "4-bit MatMulNBits, block 32, float zero point; int8 embedding; fp32 heads",
"executionProviders": [
"wasm"
],
"argmaxAgreementVsFp32": 0.981,
"note": "Smaller, and needs float zero points: at 4 bits an integer zero point collapses this checkpoint, because only a float one reproduces each block's extremes exactly. onnxruntime-web's WebGPU MatMulNBits kernels have no float zero-point path, so every matmul falls back to the CPU and it runs about 15x slower than q8 there."
}
]
}