IBM Granite Embedding Reranker English R2

Granite-Embedding-Reranker-English-R2 is the second stage of a search system. It cannot search. Given a query and a candidate document it reads both together and returns a single number — how relevant one is to the other — and that is all it does. A retrieval model fetches twenty plausible documents in milliseconds; this model then re-reads all twenty properly and decides which three deserve to be at the top. The split exists because the two jobs have incompatible economics. An embedding model encodes a query and a document separately, so a million documents can be indexed once and searched instantly, at the cost of never comparing the two texts directly. A cross-encoder like this one reads query and document as a single sequence and therefore judges relevance far more accurately, but it has to run once per candidate pair, which makes it hopeless for searching a corpus and ideal for polishing a shortlist. On IBM's own numbers, reranking the top twenty results from granite-embedding-english-r2 lifts BEIR retrieval from 53.1 to 55.8, long-document search from 41.6 to 45.8 and multilingual MIRACL from 43.6 to 55.2 — that last one a gain of nearly twelve points, from one extra pass over documents the search engine had already retrieved. IBM's comparison table also shows the model losing. Against gte-reranker-modernbert-base, a competitor of exactly the same size and context length, Granite trails on two of three benchmarks — 55.8 against 56.1 on BEIR and, more starkly, 45.8 against 51.2 on long documents. A five-point deficit on long-document reranking is the kind of number most vendors leave out of the card. IBM printed it, in bold, next to its own result. The architecture matches its retrieval sibling: ModernBERT, 149 million parameters, 22 layers, 12 attention heads, an intermediate size of 1,152, a 50,368-token vocabulary and an 8,192-token context — eight times what the older 512-token rerankers accept, which is the whole reason it holds up on long documents at all. Training used ranking-specific losses, including pListMLE, plus model merging. As across the whole R2 family, MS MARCO was excluded from training because its licence forbids commercial use — a deliberate accuracy sacrifice IBM documents rather than hides. Licence is plain Apache 2.0, no territorial carve-out. English only, and the name is misleading: despite sitting in the Granite Embedding collection, this model produces no embeddings you can store.

▸News

IBM printed a rival's better score in its own model card — and it is worth reading why

9/1/2026 · IBM, karta modelu granite-embedding-reranker-english-r2 (Hugging Face)

▸Videos

No videos yet.

▸Reviews