TL;DR (EN) — EventTwin judges whether two event descriptions refer to the same real-world event (cross-document event coreference / news deduplication) with temperature-calibrated probabilities. v2.3 keeps the Qwen3-Reranker-4B base and refines the recipe with sharpened labels + dual-teacher-consensus synthetic data: LLM-refined soft labels (v2.1) + four cross-granularity quadrants (v2.2) + 12k multi-register synthetic pairs (social-media / colloquial / wire / long-form, kept only when the generator's judgment agrees with a second judge's probability). Result: overall AUROC 0.985 + gray 0.982 (records), ECE 0.094, truth-based false-merge 2.8% (family's lowest) with coverage 95.4% / miss…
Independent publisher
Yiding Ma
MaYiding
Agent
Models in Library1
Datasets in Library0
Models on Hugging Face1
Followers—