WRSWorld Research StateContribute
WORLD RESEARCH STATE

BGE/C-Pack and M3: embedding resources and self-distilled hybrid retrieval

Contribution grade

Multilingual text retrieval · Human-designed research method or resource

Reference score120,000Provisional editorial estimate of this exact contribution; its grade is assessed separately.

What changed

Packaged Chinese evaluation/training resources and a staged embedding recipe, then extended the family with shared dense, lexical and multi-vector retrieval heads whose integrated scores teach each head through self-distillation.

Scientific scope

Packaged Chinese evaluation/training resources and a staged embedding recipe, then extended the family with shared dense, lexical and multi-vector retrieval heads whose integrated scores teach each head through self-distillation. C-Pack original C-TEM resource names and later BGE-M3 extension are source-specific stages, not retroactively identical models. M3's integrated teacher is an empirical training choice, not guaranteed more accurate than each head. Dense, sparse and late-interaction components are inherited; full multi-vector scoring has extra cost. The retrieval method does not itself establish a medical benefit.

Evidence

The linked original sources describe the stated contribution and its evidence limits. WRS has not independently rerun the experiments or certified a proof. C-Pack original C-TEM resource names and later BGE-M3 extension are source-specific stages, not retroactively identical models. M3's integrated teacher is an empirical training choice, not guaranteed more accurate than each head. Dense, sparse and late-interaction components are inherited; full multi-vector scoring has extra cost. The retrieval method does not itself establish a medical benefit.

Contributors

Provisional WRS assessment, edition 2026-10-16.1. Independent expert calibration is pending. This grade does not certify correctness or assign ownership.

Full scientific record · Browse contribution grades · Request a correction