teinam/KURE-v2-onnx

KURE-v2 (ONNX, fp32)

ONNX export of nlpai-lab/KURE-v2 (commit 3431f86d399d666083890dbb882aced6708873bc) as a ColBERT / PyLate late-interaction encoder. Used by the Ieum desktop app for on-device knowledge search.

This is a format conversion only — no fine-tuning. License follows the base model (Apache-2.0). All credit for the model goes to NLP&AI Lab, Korea University.

Files

File	Bytes	SHA256
model.onnx	614,804,784	4971279166197764dc20d14a14b5ff184489114ede9da5bdcbf57f2507dceaad
tokenizer.json	1,087,863	b9cd3dd330a256f0b73a5d93e066d58d54e283ea88eb0e96d62c2ee53c28563d
kure_meta.json	1,637	token ids, query/document rules, file hashes

ONNX Runtime (bundled for convenience)

onnxruntime/libonnxruntime.dylib is the unmodified lib/libonnxruntime.1.31.0.dylib from the official ONNX Runtime 1.31.0 osx-arm64 release (tgz SHA256 1031b36dd3bdaa8d976a6130d9cf2a05e9be320b93ae804f8427b91dc3432afc), signed by Microsoft (TeamIdentifier UBF8T346G9). MIT License — see onnxruntime/LICENSE.

File	Bytes	SHA256
onnxruntime/libonnxruntime.dylib	44,830,288	2a50a63e7f41f8d4d551d3ee47248df3864f6540109fce8ad520ca3b51af4b70

Graph

• Inputs: input_ids, attention_mask — int64, [batch, seq]
• Output: embeddings — float32, [batch, seq, 128], already L2-normalized
• ModernBERT encoder → PyLate Dense 768→1536 (residual) → 1536→768 (residual) → 768→128
• opset 18, TorchScript exporter (torch 2.8.0, transformers 5.3.0, pylate 1.6.0)
• Verified against PyLate fp32: min token cosine 0.9999999, MaxSim sum diff ≤ 1.1e-4
• Tested with ONNX Runtime 1.31.0 (CPU EP). fp16 conversion does not load on ORT 1.31 CPU, so this is fp32.

Tokenization (PyLate 1.6 rules)

• Query: tokenize with special tokens, truncate to 63, pad to 63 with <mask> (id 4, attention 0), insert [Q] (id 50000) at position 1 → 64 tokens; use all 64 vectors.
• Document: insert [D] (id 50001) at position 1; drop vectors where attention is 0 or id is in skiplist_ids.
• tokenizer.json ships with truncation/padding enabled — disable both before use.
• Score: MaxSim.

Exported with tools/export_kure_v2.py from the Ieum project.

---

KURE-v2 를 PyLate(ColBERT) 방식 그대로 ONNX(fp32)로 내보낸 것입니다. 이음(Ieum) 앱이 지식 검색 모델로 내려받아 씁니다. 형식만 바꿨고 라이선스는 원본(Apache-2.0)을 따릅니다.