mteb/Vidore3RetrievalPredictions

ViDoRe v3 Retrieval Predictions

Precomputed first-stage retrieval predictions for ViDoRe v3 tasks, for reuse as reranking candidates. This repository stores prediction files; corpus documents, queries and relevance judgments remain in the source task datasets.

File layout

<model-organisation>__<model-name>/<document-representation>/<task-name>_predictions.json

Model directories follow MTEB's model_name_as_path() convention: replace / with __ and spaces with _, preserving case. For example, Qwen/Qwen3-VL-Embedding-2B becomes Qwen__Qwen3-VL-Embedding-2B. The original model ID and revision remain unchanged inside each JSON file. Representation directories describe the actual document inputs used by that retrieval run:

• text: document text only.
• image: document images only.
• text-image: document text and images together.

Each file retains MTEB's original JSON structure and bytes: mteb_model_meta records the producer model and revision, and language subset / evaluation split / query ID / document ID mappings contain retrieval scores. Task versions remain in filenames. Languages actually evaluated are those present in each file.

Use a pinned Hugging Face dataset commit when referencing a prediction file for evaluation. Available candidate depth must cover the requested reranking depth.

Prepared predictions

Each model has prediction files for all eight domains: ComputerScience, Energy, FinanceEn, FinanceFr, Hr, Industrial, Pharmaceuticals and Physics.

Retriever	Document representation	Directory
mteb/baseline-bm25s	Text	mteb__baseline-bm25s/text/
BAAI/bge-m3	Text	BAAI__bge-m3/text/
Qwen/Qwen3-VL-Embedding-2B	Text and images	Qwen__Qwen3-VL-Embedding-2B/text-image/

The representation describes the first-stage retrieval run, independently of the document inputs used by a subsequent reranker.