Jourdain/CoRAFIG-phase1

CoRAFIG — Phase 1 (Annotator Calibration)

CoRAFIG (Corpus de Résumés Automatiques Francophones Inter-Genres) is a French human-annotated corpus for evaluating automatic summarization quality across multiple domains. It is developed in the framework of the RAFFAL project (DGA / AMU / ChapsVision / AgroParisTech), inspired by the SummEval benchmark.

This dataset contains Phase 1, which served as the annotator calibration phase: 20 source documents were summarized by 10 automatic systems and evaluated by expert annotators to align annotation practices before the main annotation campaign.

Dataset at a Glance

Documents	15
Summaries	120 (15 docs × 8 systems on average)
Annotations	360
Annotators	6 (expert native speakers)
Systems	10 (see below)
Language	French
Domain	Multi-genre (news, legal, financial, research, audio)

Annotation Dimensions

Each summary is rated on a 1–5 scale across 5 quality criteria:

Criterion	Description
quality_criterion_1_fluency	Grammatical and syntactic correctness
quality_criterion_2_readability	Ease of reading and natural flow
quality_criterion_3_coherence	Internal logical consistency of the summary
quality_criterion_4_consistency-factuality	Factual alignment with the source document
quality_criterion_5_relevance	Coverage of the key information from the source

In addition, annotators tagged error spans within the summary text with one of three categories:

• lang_errors — linguistic errors (grammar, spelling, agreement, etc.)
• fact_errors — factual inconsistencies with the source
• source_reference — references that are unsupported by or contradictory to the source

Automatic Systems

10 summarization systems were evaluated:

System	Type
gpt-4o	LLM (OpenAI)
mistral-small	LLM (Mistral AI)
qwen3-32b	LLM (Alibaba)
qwen3-8b	LLM (Alibaba)
barthez	Fine-tuned encoder-decoder (French BART)
mbart	Fine-tuned multilingual BART
text_rank	Extractive (TextRank)
clust	Extractive (clustering-based)
ner	Extractive (NER-guided)
ner_clust	Extractive (NER + clustering)

Note: Phase 1 includes barthez and mbart which were dropped in Phase 2.

Data Structure

The dataset follows a nested JSON hierarchy:

documents[]
  └── document_id, source_text
  └── summaries[]
        └── summary_id, technology, summary_text
        └── annotations[]
              └── annotator_id, annotator_level, annotation_id
              └── ratings: {criterion_1..5}
              └── labels: [{category, start, end, text, labels}]
              └── notes: {source_reading_time, summary_reading_time, annotation_time}

Usage

import json
from datasets import load_dataset

# Option A — raw JSON
with open("data/delivery_enriched--20260403-131206.json") as f:
    data = json.load(f)

for doc in data["documents"]:
    source = doc["source_text"]
    for summary in doc["summaries"]:
        system = summary["technology"]
        for ann in summary["annotations"]:
            ratings = ann["ratings"]
            labels = ann["labels"]  # error spans

Relation to Phase 2

This Phase 1 dataset is the calibration set. The main annotation campaign is CoRAFIG Phase 2, targeting ~3,000 annotations across 100 documents.

Source Corpora

Documents were drawn from five existing French summarization corpora:

• News: MLSUM (Scialom et al., EMNLP 2020)
• Legal: EUR-Lex-Sum (Aumiller et al., EMNLP 2022)
• Financial: CoFiF-Plus (Moro et al., TSAR 2022)
• Research: HAL open archive (hal.science)
• Audio: FREDSum (Kante et al., EACL 2023)

License

Creative Commons Attribution 4.0 International (CC BY 4.0)