---
license: cc-by-nc-4.0
language: en
pretty_name: CiteRight RAG Corpus (IP and Innovation Policy)
tags:
- rag
- retrieval
- citation
- faithfulness
- scientific-papers
size_categories:
- 1K<n<10K
---

# CiteRight Corpus

Research dataset accompanying the CS-552 (EPFL Spring 2026) Open Project
**Faithful RAG: Citation Accuracy and Retrieval Robustness in Domain-Specific
Scientific Literature** by Team CiteRight (Elie Bruno, Andrea Trugenberger,
Faruk Zahiragic, Yusif Askari).

## Contents

| Path | Description |
|---|---|
| `chunks/<paper_id>_<coarse\|fine>.json` | Pre-computed semantic chunks per paper (~2 000 chars coarse, ~300 chars fine). |
| `indexes/<coarse\|fine>.faiss` | FAISS `IndexFlatIP` built from OpenAI `text-embedding-3-small` (1 536-d). |
| `indexes/<coarse\|fine>_metadata.json` | Row → chunk mapping with paper id, section hierarchy, text. |
| `processed/<paper_id>/document.md` | Dolphin-1.5 parsed markdown (preserves sections, equations, tables). |
| `processed/<paper_id>/metadata.json` | Page-level layout metadata from the VLM. |
| `raw_metadata/<paper_id>.json` | OpenAlex bibliographic record (DOI, year, OA status). |

Original PDFs are **not** redistributed: anyone who needs them can refetch
from OpenAlex using the `id` field in `raw_metadata/<paper_id>.json`.

## Provenance

- **Source:** OpenAlex, filtered with `primary_topic.id=t10856` and `open_access.is_oa=true`.
- **PDF parsing:** [Dolphin 1.5](https://github.com/ByteDance-Seed/Dolphin) on a CUDA 11.8 worker fleet.
- **Embeddings:** OpenAI `text-embedding-3-small` (1 536-d), normalised for cosine similarity.

## Loading

```python
from huggingface_hub import snapshot_download

local = snapshot_download("citeright/corpus", repo_type="dataset",
                          local_dir="/scratch/citeright_artifacts")
```

Or use the helpers in [`evaluation/common/data_loader.py`](https://github.com/eliemada/cs552-faithful-rag/blob/main/evaluation/common/data_loader.py)
which auto-resolves to the right cache.

## License

Released under [CC-BY-NC-4.0](https://creativecommons.org/licenses/by-nc/4.0/).
Each underlying paper retains its own license. For takedown requests, open
a discussion on this repo.

## Citation

```bibtex
@misc{citeright2026,
  title  = {CiteRight: Citation Accuracy and Retrieval Robustness in Scientific RAG},
  author = {Bruno, Elie and Trugenberger, Andrea and Zahiragic, Faruk and Askari, Yusif},
  year   = 2026,
  note   = {CS-552 Modern NLP Open Project, EPFL}
}
```