elcodabra/seklu-reranker-bge-m3-coreml

bge-reranker-v2-m3, Core ML (int8)

BAAI/bge-reranker-v2-m3 converted to Core ML for the optional "thorough search" of Seklu / Finder+ on macOS: it reorders the 20 files nearest to a query by reading each with it.

reranker-bge-m3-coreml-v1.zip (520,748,750 bytes, SHA-256 d7cc20484cff7bd8a42f2464ffb8dbbbd2bf6b6f766988c898caa99d42c018d7) holds BAAI.bge-reranker-v2-m3/ with:

• Reranker.mlmodelc: inputs input_ids, attention_mask (int32, 8×256: eight <s> query </s></s> passage </s> pairs, pad 1), output score (8 logits, higher = more relevant). int8 weights, fp16 compute, macOS 15+. Run it on CPU + GPU: on the Neural Engine part of the graph fails to compile and every load takes 40-100 s.
• vocab.tsv: the XLM-R SentencePiece vocabulary (piece TAB score, line = id), the same pieces and ids as multilingual-e5's.

Checked: logits within 0.03 of the Python Core ML run and 0.13 of PyTorch; 0.19 s per 8 pairs on an Apple silicon GPU. On a private set of 44 file-search queries over real personal documents, reordering the top 20 files of the e5 v3 model: hit@1 20 → 25, hit@10 39 → 42, MRR 0.64 → 0.71.

License: Apache-2.0, as the original model.