ELO-GRAG

Joint EURECOM-Lettria-Orange submission workspace for the EvalLLM 2026 RAG Challenge.

Layout

challenge/      Challenge output schemas, sample queries, and validation.
data/           Local datasets (git-ignored).
experiments/    Individual experiments.
runs/           Generated outputs, logs, and notes (git-ignored).
submissions/    Final JSON files for submission.
tools/          Shared helper utilities.

Local Data

Expected local data layout:

ELO-GRAG/
  data/
    Corpus_raw/
    Corpus_parsed/

Download the dataset archives manually:

• Corpus_raw.tar.gz: https://nextcloud.eurecom.fr/s/4t4bKTjqo6FiS7o
• Corpus_parsed.tar.gz: https://nextcloud.eurecom.fr/s/aKDHteYdLwHBSqb

Place both archives in data/, then extract them:

mkdir -p data
tar -xzf data/Corpus_raw.tar.gz -C data
tar -xzf data/Corpus_parsed.tar.gz -C data

• Corpus_raw/ contains the original PDFs.
• Corpus_parsed/ contains one subfolder per document, with the output from docparsing (JSON, MD, XML formats) based on the original PDFs, along with a source.txt file containing the original PDF filename for reference.

Experiments

See experiments/README.md.

Answer Generation

See tools/answering/README.md.