What is this?
ONNX export (opset 18, dynamic batch and sequence) of the Mira-v1 encoder,
plus heads.pt scorer weights and the tokenizer. Verified numerically
against PyTorch (max abs diff 2.9e-5 — fp32 noise).
Use
import numpy as np, onnxruntime as ort
from transformers import AutoTokenizer
tok = AutoTokenizer.from_pretrained("sagea-ai/Mirav1_onnx")
sess = ort.InferenceSession("model.onnx", providers=["CPUExecutionProvider"])
b = tok(["My payouts are failing. [SEP] billing"], return_tensors="pt")
h = sess.run(None, {"input_ids": b["input_ids"].numpy(),
"attention_mask": b["attention_mask"].numpy()})[0][:, 0]
Score with heads.pt, softmax, done. See the main card for benchmarks,
limitations, and license (Apache-2.0).
Playground
Try it live, no install: Mira Playground Space (Gradio + ONNX CPU; WebGPU in-browser build tracked as follow-up).