README.md

data Directory

Holds the dataset and generated artifacts. File contents are git-ignored; folder structure is tracked.

Table of Contents

• 1. Purpose
• 2. Subdirectories
• 3. Expected Layout After Running The Pipeline
• 4. Notes

1. Purpose

This directory stores the BEIR SciFact dataset and the outputs of the pipeline. The actual files (datasets, reports) are intentionally not committed to git (see the root .gitignore). Only the folder structure is kept, via .gitkeep files and this README.

2. Subdirectories

Folder	Contents	Produced by
raw/	The dataset exactly as downloaded from HuggingFace: corpus, queries, and qrels.	make download-data (Phase 1)
processed/	Normalized files used by the pipeline: corpus.jsonl, queries.jsonl, qrels.jsonl.	make prepare-data (Phase 2)
reports/	Generated benchmark output, e.g. retrieval_benchmark.md.	make evaluate (Phase 8)

3. Expected Layout After Running The Pipeline

data/
├── raw/
│   ├── corpus/        # downloaded corpus documents
│   ├── queries/       # downloaded queries
│   └── qrels/         # downloaded relevance labels
├── processed/
│   ├── corpus.jsonl   # {id, title, text, ...} per line
│   ├── queries.jsonl  # {id, text} per line
│   └── qrels.jsonl    # {query_id, doc_id, relevance} per line
└── reports/
    └── retrieval_benchmark.md

4. Notes

• The exact schema of the processed files is defined by the data preparation code in src/rag_retrieval_benchmark/data/ and documented in that folder's README.
• Because these files can be large, regenerate them locally with the Make commands rather than committing them.