Kostya165/LFM2.5-Encoder-350M-PII-Detector-ONNX

LFM2.5-Encoder-350M-PII-Detector ONNX

FP32 ONNX export of LiquidAI/LFM2.5-Encoder-350M-PII-Detector: the bidirectional encoder plus the BIOES token-classification head (161 labels, 40 PII types). Opset 18. Weights are one file, model.onnx (about 1.4 GB).

The graph is a single forward without a KV cache.

inputs	input_ids int64 [batch, sequence], attention_mask int64 [batch, sequence]
output	logits float32 [batch, sequence, 161]
batch, sequence	dynamic

Against the PyTorch model (attn_implementation="eager", fp32) the largest absolute difference on checked shapes was 2.6e-5, and argmax labels matched, including padded batches and length 1.

Labels are BIOES. id2label / label2id are in config.json. O is outside an entity; B / I / E are begin, inside, end of a multi-token span; S is a single-token span.

Usage

import json
import numpy as np
import onnxruntime as ort
from transformers import AutoTokenizer

repo = "Kostya165/LFM2.5-Encoder-350M-PII-Detector-ONNX"
tokenizer = AutoTokenizer.from_pretrained(repo)
labels = {int(k): v for k, v in json.load(open("config.json"))["id2label"].items()}

session = ort.InferenceSession("model.onnx", providers=["CPUExecutionProvider"])
encoded = tokenizer("Email ada@example.com or call +1 415 555 0134.", return_tensors="np")
logits = session.run(None, {
    "input_ids": encoded["input_ids"],
    "attention_mask": encoded["attention_mask"],
})[0]
pred = logits.argmax(-1)[0]
for token, label_id in zip(tokenizer.convert_ids_to_tokens(encoded["input_ids"][0]), pred.tolist()):
    label = labels[label_id]
    if label != "O":
        print(token, label)

License

Derivative of Liquid AI's model. Use, reproduction, and distribution follow LFM Open License v1.0.