---
dataset_info:
- config_name: corpus
  features:
  - name: _id
    dtype: string
  - name: title
    dtype: string
  - name: text
    dtype: string
  - name: metadata
    struct:
    - name: document_id
      dtype: string
    - name: page_number
      dtype: int64
  - name: image
    dtype: 'null'
  splits:
  - name: test
    num_bytes: 43537055
    num_examples: 22700
  download_size: 17741057
  dataset_size: 43537055
- config_name: corpus-with-image
  features:
  - name: _id
    dtype: string
  - name: title
    dtype: string
  - name: text
    dtype: string
  - name: metadata
    struct:
    - name: document_id
      dtype: string
    - name: page_number
      dtype: int64
  - name: image
    dtype: image
  splits:
  - name: test
    num_bytes: 2295554066.0
    num_examples: 22700
  download_size: 2267597419
  dataset_size: 2295554066.0
- config_name: queries
  features:
  - name: _id
    dtype: string
  - name: query_text
    dtype: string
  - name: relevant_document_ids
    list:
    - name: corpus_id
      dtype: string
    - name: metadata
      struct:
      - name: document_id
        dtype: string
      - name: page_number
        dtype: int64
      - name: question_type
        dtype: string
    - name: score
      dtype: float64
  splits:
  - name: test
    num_bytes: 16544
    num_examples: 67
  download_size: 9270
  dataset_size: 16544
configs:
- config_name: corpus
  data_files:
  - split: test
    path: corpus/test-*
- config_name: corpus-with-image
  data_files:
  - split: test
    path: corpus-with-image/test-*
- config_name: queries
  data_files:
  - split: test
    path: queries/test-*
---

# ECSS-1.0 Dataset

## Dataset Summary

This dataset provides a focused benchmark for retrieval and generation tasks related to ECSS (European Cooperation for Space Standardization) documents. It includes a set of documents, queries, relevance judgments (qrels), and page images.

- Number of Documents: 196
- Number of Queries: 67
- Number of Pages: 22700
- Number of Relevance Judgments (qrels): 89
- Average Number of Pages per Query: 1.3

## Dataset Structure (Hugging Face Datasets)

The dataset is structured into the following subsets:

- `corpus`: Contains page-level information:
  - `_id`: A unique identifier for this specific page within the corpus.
  - `title`: The title of the document.
  - `text`: The text of the document.
- `queries`: Contains query information:
  - `_id`: Unique identifier for the question.
  - `query_text`: The question text.
  - `relevant_document_ids`: A list of corpus documents considered as references for this question, each reference containing:
    - `corpus_id`: The document identifier.
    - `score`: The importance or relevance score.

## Usage Examples

You can load the datasets using the `load_from_disk` function from the `datasets` library. Replace the paths with the actual locations on your machine.

```python
from datasets import load_dataset
dataset_queries_test = load_dataset("FOR-sight-ai/ECSS-1.0", "queries", split="test")
```

## Results

| Model Name | nDCG@10 |
| :--- | ---: |
| bm25 | 0.43 |
| bge-large-en-v1.5 | 0.44 |
| nomic-embed-multimodal-3b | 0.59 |
| colqwen2.5-v0.2 | 0.68 |

## Citation

If you use this dataset in your research or work, please cite:

```bibtex
@misc{ecssbenchmark2025,
  title={ECSS RAG benchmark},
  author={Francois Lancelot and Nawal Ould Amer and Benjamin Fourreau and Catherine Kobus and Marion-Cécile Martin},
  primaryClass={cs.IR},
  year={2025},
}
```