README.md

---
pretty_name: SupraReviewBench
language:
- en
task_categories:
- text-generation
- question-answering
size_categories:
- 1K<n<10K
license: cc-by-4.0
configs:
- config_name: default
  data_files:
  - split: train
    path: benchmark.jsonl
---

# SupraReviewBench

## Dataset summary
SupraReviewBench is a peer-review benchmark built from OpenReview discussion threads.
Each record represents one paper and its full review discussion. Reviewer opinions are
split into atomic blocks, labeled with a taxonomy, grouped by discussion point, and
validated for correctness via conflict adjudication and author-refutation analysis.

The dataset is intended for opinion-level evaluation and training, with explicit
labels that mark which reviewer opinions are likely correct or incorrect.

## Dataset viewer
The Dataset Viewer reads `benchmark/benchmark.jsonl`. The YAML config above
declares a single `train` split and an explicit schema so the Viewer renders
columns normally instead of wrapping records into a single text field.

## Source and coverage
- Source: OpenReview discussions (ICLR and NeurIPS). Use the `conference` field
  in each record to see the exact venue and year.
- Unit: one paper (OpenReview forum id).
- Language: primarily English.

## Data format and fields
The dataset is stored as a JSONL file named `benchmark/benchmark.jsonl`.
Each JSONL line is a single JSON object with the following top-level fields.
Some fields are optional depending on the paper or venue.

Core fields:
- `id`: OpenReview forum id for the paper (string, unique).
- `conference`: venue label (e.g., "ICLR 2017").
- `content`: paper metadata from OpenReview (title, abstract, authors, pdf path, etc.).
- `decision`: acceptance decision string.
- `reviews`: review discussion threads, each a list of `[role, payload]` pairs.
- `metareview`: meta-review threads with the same `[role, payload]` structure.
- `sentence_texts`: list of atomic sentences; indices are referenced elsewhere.
- `opinions`: list of labeled opinion blocks (see below).
- `opinion_groups`: list of groups; each group is a list of opinion indices that
  discuss the same point.
- `conflicts_validation`: list of "correct"/"incorrect" labels aligned to `opinions`.
- `rebuttal_validation`: list of "correct"/"incorrect" labels aligned to `opinions`.

Opinion block structure:
Each entry in `opinions` is a 2-element list:
1) `sources`: list of `[role, [sentence_ids]]` pairs
2) `tags`: list of taxonomy labels (multi-label)

Example (simplified):
```json
{
  "id": "rk9eAFcxg",
  "conference": "ICLR 2017",
  "opinions": [
    [
      [["Reviewer 1", [0, 1]], ["Author", [4, 5]]],
      ["QUAL-EXP", "QUAL-CMP"]
    ]
  ],
  "opinion_groups": [[0]],
  "conflicts_validation": ["correct"],
  "rebuttal_validation": ["correct"],
  "PDF_path": "benchmark/PDF/ICLR2017_rk9eAFcxg.pdf",
  "MD_path": "benchmark/MD/ICLR2017_rk9eAFcxg.md"
}
```

## Taxonomy labels
Labels follow a fixed taxonomy with 5 coarse categories and sublabels:
- QUAL (Quality): QUAL-MET, QUAL-EXP, QUAL-REP, QUAL-CMP, QUAL-STA
- CLAR (Clarity): CLAR-WRT, CLAR-NOT, CLAR-FIG
- SIGN (Significance): SIGN-BRD, SIGN-DOM, SIGN-SOT, SIGN-IMP
- ORIG (Originality): ORIG-PROB, ORIG-MTH, ORIG-ANL, ORIG-EXP, ORIG-COM, ORIG-NEG
- POL (Policy/Compliance): POL-ETH, POL-DAT, POL-ANO, POL-PLG, POL-IMP
- N/A: polite text or non-substantive content

## Annotations and validation
Two validation signals are provided, each aligned to `opinions`:
- `conflicts_validation`: results of reviewer opinion conflict adjudication.
- `rebuttal_validation`: results of author refutation validation.

Values are "correct" or "incorrect".

## PDF and Markdown files
`PDF_path` and `MD_path` are string paths to local assets used during curation.
These files are not included in the dataset repo (PDFs are too large). The fields
remain as strings and do not affect Dataset Viewer loading.

## Intended use
This dataset is designed for:
- multi-label classification of reviewer opinions
- opinion grouping and conflict detection
- evaluation of reviewer correctness and disagreement

It is not intended for ranking papers or making accept/reject decisions.

## Limitations
- Labels are produced with LLM assistance and are not perfect.
- Some venues and years may have missing or incomplete review metadata.
- PDF and Markdown assets are not included in the dataset repo.

## License
This dataset is released under CC BY 4.0.

## Citation
If you use this dataset, please cite the associated paper or this repository.
Add a BibTeX entry here if you have a preferred citation.