[Ryenhails]

This PR adds results of EVIE-4.5B + JEV-27B-VL to the ViDoRe V3 pipeline benchmark.

EVIE-4.5B retrieves the top 100 pages. JEV-27B-VL reranks the top 16 page images with one multiple-choice decision per query. It uses one forward pass and generates no text.

The average nDCG@10 over the 8 English datasets is 0.6825. Search takes 1.05 s per query with 8 concurrent queries on one H200.

The CLI writes num_queries only per language, so we copied the English value into overall for the CI check. The pipeline implementation is available at Jev-reranker.

[QuentinJGMace]

Hey @Ryenhails, thanks for this one too, it looks great! Same kind of minor comments as on #145:

• You ran evals with num_concurrent=8, so the 1.05 s/query is throughput: each query spends ~8.4 s in the reranker. Other entries ran sequentially, so this isn't comparable on the leaderboard. Could you run a short sequential timing (e.g. 30–50 queries)? Otherwise I'll add a footnote.
• The reranker prompt includes each page's doc_id (Page A (document: <doc_id>)), which is often a descriptive title.
• Could you add an info_url to your README for clarity?

[Ryenhails]

Thanks!

You ran evals with num_concurrent=8, so the 1.05 s/query is throughput: each query spends ~8.4 s in the reranker. Other entries ran sequentially, so this isn't comparable on the leaderboard. Could you run a short sequential timing?

We ran the first 40 queries of hr with num_concurrent=1 on the same H200, with the submitted pipeline file. Search takes 1.47 s per query. We added this to the hardware notes.

The reranker prompt includes each page's doc_id (Page A (document: <doc_id>)), which is often a descriptive title.

Yes. The description now mentions it.

Could you add an info_url to your README for clarity?

Done. info_url points to the README.

[QuentinJGMace]

LGTM, merging