---
pretty_name: RAVENEA for MTEB
task_categories:
- image-to-text
dataset_info:
- config_name: corpus
  features:
  - name: id
    dtype: string
  - name: text
    dtype: string
  splits:
  - name: test
    num_bytes: 135936508
    num_examples: 11396
  download_size: 71264021
  dataset_size: 135936508
- config_name: default
  features:
  - name: query-id
    dtype: string
  - name: corpus-id
    dtype: string
  - name: score
    dtype: int64
  - name: culture_relevance
    dtype: int64
  splits:
  - name: test
    num_bytes: 145682
    num_examples: 1610
  download_size: 20744
  dataset_size: 145682
- config_name: queries
  features:
  - name: id
    dtype: string
  - name: image
    dtype: image
  - name: country
    dtype: string
  - name: category
    dtype: string
  - name: task_type
    dtype: string
  splits:
  - name: test
    num_bytes: 67375866
    num_examples: 161
  download_size: 67375484
  dataset_size: 67375866
- config_name: top_ranked
  features:
  - name: query-id
    dtype: string
  - name: corpus-ids
    list: string
  splits:
  - name: test
    num_bytes: 39143
    num_examples: 161
  download_size: 33603
  dataset_size: 39143
configs:
- config_name: corpus
  data_files:
  - split: test
    path: corpus/test-*
- config_name: default
  data_files:
  - split: test
    path: data/test-*
- config_name: queries
  data_files:
  - split: test
    path: queries/test-*
- config_name: top_ranked
  data_files:
  - split: test
    path: top_ranked/test-*
---

# RAVENEA for MTEB

This repository is a deterministic MTEB normalization of
[`jaagli/ravenea`](https://huggingface.co/datasets/jaagli/ravenea) at revision
`ff5a212a9bfa1515f82e0930b37b7d64e3e9ee2e`. It contains the official 161-image test split, the complete
11,396-document Wikipedia corpus, 1,610 human judgments, and each query's ten
official retrieval candidates.

The `culture_relevance` column preserves the source grade from -3 to 3. The
qrels `score` column stores `2 ** (culture_relevance + 3) - 1`, exactly matching
the gain used by the official RAVENEA nDCG implementation.

The source dataset card does not declare a dataset-wide license. The images are
adapted from CVQA and CCUB, and the corpus contains Wikipedia-derived text.
Users should review the upstream dataset and source-item terms before reuse.

Generated by `scripts/data/ravenea/create_data.py` in MTEB. The source archive
SHA-256 is `4ef30a2c35ce2ebee39f05128e5a1042f562158ff4a2929d835af3e5f20772bc`.