CloveAI/clov-embed-v2

import onnxruntime as ort
import numpy as np
from transformers import AutoTokenizer
from huggingface_hub import hf_hub_download

# ── Load ───────────────────────────────────────────────────────────────────
tokenizer    = AutoTokenizer.from_pretrained("CloveAI/clov-embed-v2", subfolder="tokenizer")
onnx_path    = hf_hub_download("CloveAI/clov-embed-v2", "onnx/biencoder_rope.onnx")
session      = ort.InferenceSession(onnx_path, providers=["CPUExecutionProvider"])

# ── Encode ─────────────────────────────────────────────────────────────────
def encode(texts):
    if isinstance(texts, str): texts = [texts]
    enc = tokenizer(texts, padding=True, truncation=True, max_length=256, return_tensors="np")
    return session.run(["embeddings"], {"input_ids": enc["input_ids"], "attention_mask": enc["attention_mask"]})[0]

# ── Test ───────────────────────────────────────────────────────────────────
emb = encode("Hello world!")
print(emb)   # (1, 256)

BiEncoder RoPE — Sentence Embedding Model

A 34M parameter sentence embedding model trained from scratch using PyTorch.

Architecture

• 6-layer Transformer encoder with RoPE positional embeddings
• Mean pooling + L2 normalization
• 256-dim output vectors

Training (Curriculum)

Phase	Dataset	Loss
1	all-nli	MNRLoss
2	squad	MNRLoss
3	msmarco-bm25	HardNegativeLoss
4	natural-questions	MNRLoss

Files

• tokenizer/ — HuggingFace tokenizer (bert-base-uncased)
• pytorch/checkpoint_phase4_nq.pt — PyTorch weights
• onnx/biencoder_rope.onnx — ONNX FP32
• onnx/biencoder_rope_int8.onnx — ONNX INT8 (recommended for CPU)

Performance

• FP32 ONNX size : 134.3 MB
• INT8 ONNX size : 34.6 MB
• Throughput : ~247 sentences/sec on CPU