bge-reranker-v2-m3 — ONNX

ONNX (dynamic int8 quantization, q8) conversion of BAAI/bge-reranker-v2-m3 (revision 953dc6f6f85a1b2dbfca4c34a2796e7dde08d41e) for on-device reranking in the PLaiA app (knowledge search: query–passage relevance score).

• Original model: BAAI (Beijing Academy of Artificial Intelligence), FlagEmbedding — Apache License 2.0.
• Changes (Modified by PiterQ): format conversion only (ONNX (dynamic int8 quantization, q8)); input fixed to input_ids, attention_mask [1, 512] (int32 in Core ML, int64 in ONNX); output score (logit; sigmoid → 0–1). No fine-tuning, no change of training data.
• Tokenizer: the original XLM-RoBERTa tokenizer files of the source model, unchanged.
• meta.json: token ids and shape used by the app.

License: Apache License 2.0 (see LICENSE and NOTICE). Citation: Chen et al., BGE M3-Embedding (2024), and the FlagEmbedding repository.