--- license: cc-by-4.0 task_categories: - text-generation - text-classification - summarization - question-answering language: - en tags: - peer-review - openreview - scientific-papers - nlp - benchmarking - meta-science pretty_name: "ReviewArena: A Large-Scale Cross-Conference Dataset and Benchmark for LLM Peer Review" size_categories: - 10K/reviewarena") print(ds) # DatasetDict({ # neurips: Dataset(num_rows=...) # iclr: Dataset(num_rows=...) # icml: Dataset(num_rows=...) # tmlr: Dataset(num_rows=...) # emnlp: Dataset(num_rows=...) # corl: Dataset(num_rows=...) # colm: Dataset(num_rows=...) # }) ``` ## Coverage One **split per conference family**. Within a split, filter by `year`, `venue_id`, or `track` for slicing. | Split | Venues / years | Tracks | ≈ Papers | |---------|--------------------------------------------------------------------------------------|-----------------------------------------|----------| | neurips | 2021, 2022, 2023, 2023 D&B, 2024, 2025 | main + Datasets & Benchmarks (2023) | ~18,400 | | iclr | 2020, 2021, 2022, 2023, 2024, 2025, 2026 | main | ~22,500 | | icml | 2025 | main | ~3,300 | | tmlr | rolling (all accepted papers as of April 2026) | main | ~3,750 | | emnlp | 2023 | main + Findings | ~2,000 | | corl | 2021, 2022, 2023, 2024 | main | ~820 | | colm | 2024, 2025 | main | ~720 | NeurIPS 2020 and earlier used CMT and have no public OpenReview reviews. ## Schema Each row is one paper. | Column | Type | Notes | |-------------------|-----------------|-------------------------------------------------------------------------| | `forum_id` | `string` | OpenReview forum ID — primary key | | `conference` | `string` | `neurips` / `iclr` / `icml` / `tmlr` / `emnlp` / `corl` / `colm` | | `year` | `int32` | Conference year | | `track` | `string` | `main` / `datasets_and_benchmarks` / `findings` / etc. | | `venue_id` | `string` | OpenReview venue ID, e.g. `NeurIPS.cc/2024/Conference` | | `paper_number` | `int32` | Submission number (nullable) | | `title` | `string` | | | `abstract` | `string` | | | `authors` | `list` | | | `keywords` | `list` | | | `tldr` | `string` | | | `primary_area` | `string` | | | `venue` | `string` | Final venue string, e.g. `"NeurIPS 2024 poster"` | | `decision` | `string` | `Accept (poster)`, `Accept (oral)`, `Reject`, etc. | | `decision_comment`| `string` | Area-chair meta-review | | `author_rebuttal` | `string` | General rebuttal (≤2024); empty when per-reviewer rebuttals are used | | `num_reviews` | `int32` | Convenience count | | `reviews_json` | `string` | All reviews as a JSON string — see schema below | | `markdown` | `string` | OCR'd full text of the PDF (see *OCR* below); `""` if PDF unavailable | | `markdown_chars` | `int64` | `len(markdown)` for fast filtering | ### Decoding `reviews_json` `reviews_json` is `json.dumps(list[dict])`. Decode with: ```python import json df = ds["neurips"].to_pandas() df["reviews"] = df["reviews_json"].map(json.loads) print(df.iloc[0]["reviews"][0].keys()) ``` Each review dict is a **union over all forms used by all venues across all years**, with absent fields as empty strings / `None`. The most reliably populated fields are: | Field | Where populated | |-------------------|---------------------------------------------------------------| | `review_id`, `reviewer`, `rating`, `confidence`, `rebuttal` | All venues | | `summary`, `questions`, `limitations`, `strengths`, `weaknesses` | NeurIPS 2022–24, ICLR, ICML, CoRL, COLM | | `soundness`, `presentation`, `contribution` | NeurIPS 2022–24, ICLR, ICML | | `quality`, `clarity`, `significance`, `originality` | NeurIPS 2025, TMLR | | `strengths_and_weaknesses` | NeurIPS 2025 (merged form) | | `was_revised`, `final_justification` | NeurIPS 2025 (in-place review revisions) | | `claims_and_evidence`, `theoretical_claims`, `experimental_designs_or_analyses`, `relation_to_broader_scientific_literature`, `essential_references_not_discussed` | TMLR | | `paper_topic_and_main_contributions`, `reasons_to_accept`, `reasons_to_reject`, `excitement`, `reproducibility`, `ethical_concerns` | EMNLP 2023 | | `summary_of_paper`, `summary_of_recommendation`, `technical_quality`, `clarity_of_presentation`, `potential_impact`, `robotics_focus` | CoRL | | `extra_scores`, `extra_text` | dicts capturing any venue-specific fields not in the union | Schema differences in detail: - **NeurIPS 2021 D&B (track)**: older form, most numeric sub-scores absent; lives in the `neurips` split. - **NeurIPS 2022–2024**: `soundness/presentation/contribution`; separate `strengths`/`weaknesses`; single general `author_rebuttal`. - **NeurIPS 2025**: `quality/clarity/significance/originality`; merged `strengths_and_weaknesses`; per-reviewer rebuttals; in-place review revisions tracked via `was_revised` + `final_justification`. - **ICLR 2020–2026**: NeurIPS-style `soundness/presentation/contribution`; per-reviewer rebuttals. - **ICML 2025**: similar to NeurIPS-style; some venue-specific fields (e.g. `technical_quality`, `novelty`) live in `extra_scores`. - **TMLR**: claim-evidence-style structured review; rolling acceptance. - **EMNLP 2023**: ARR-style review form with `reasons_to_accept`/`reasons_to_reject`/`excitement`/`reproducibility`. - **CoRL**: robotics-focused review form. - **COLM**: language-modeling-focused review form. ## Source and collection - **Source**: [OpenReview](https://openreview.net) — main and track-level conferences (including NeurIPS Datasets & Benchmarks where applicable). - **Collected**: April 2026 via the [OpenReview Python API](https://github.com/openreview/openreview-py) (a mix of `openreview.api` v2 and legacy v1 for older NeurIPS/ICLR years). - **Pipeline**: parallel Modal workers with rate-limit-aware scheduling; each forum includes official reviews, official comments, decisions, and author rebuttals where exposed by OpenReview; PDFs ingested for OCR. ## Markdown / OCR - **Engine**: [`nvidia/nemotron-ocr-v2`](https://huggingface.co/nvidia/nemotron-ocr-v2) - **Compute**: multi-GPU Modal deployment (~6 h wall-clock end-to-end for the full OCR pass referenced in the paper) - **Throughput**: ~7,200 PDFs/hour aggregate; mean **7.3 s/PDF** per worker; **558,785** pages processed - **Failures**: 1 PDF errored during OCR; ~76 PDFs were unavailable from OpenReview at scrape time and have `markdown == ""`. Use `markdown_chars > 0` to filter. ### OCR quality (spot-checked across NeurIPS, ICLR, ICML, TMLR, CoRL, COLM) What works well: - Body prose, abstracts, section headers, paragraph structure - In-line citations like `(Author et al., YEAR)` mostly preserved - Equations rendered linearly (variable names + structure visible) - Page boundaries marked with `\n\n---\n\n` separator Recurring artifacts to be aware of (consistent across venues, low impact for most NLP tasks): - Email addresses and URLs containing repeated chars (e.g. `name@@@cmu.eed`, `https:////aaaaaaa`) - Occasional word-doubling at line breaks (`decision-decision-making`, `Complex-Valuee Valued`) - Citation lists missing semicolons (`(Singer 2007 Uhlhaas et al. 2009)`) - Greek letters and super/sub-scripts often dropped or flattened - First-page logo/header text occasionally bleeds into title (`git Cooperative …`) - Figure caption tokens interleave with body text on figure-heavy pages **Practical impact**: fine for dense retrieval, language modeling, review-grounding, summarization. **Not** suitable for tasks requiring exact equations or canonical citation strings. ## Suggested uses - Train review-quality classifiers / score predictors - Study reviewer agreement, rebuttal effectiveness, decision dynamics - Build retrieval-augmented or grounded scientific-paper assistants - Meta-research on the peer-review process across venues and years - Few-shot / RAG benchmarks that require the paper full text + the reviews ## Benchmark ## Benchmark **ReviewArena-Eval** is the companion benchmark slice: a fixed 1,002-paper subset (167 papers per numerically rated venue — NeurIPS, ICLR, ICML, CoRL, COLM, EMNLP) with venue-year-aware prompts that respect each venue's native rating scale and sub-score axes. See [`anonymousNeurIPS2026submission4281/reviewarena-eval`](https://huggingface.co/datasets/anonymousNeurIPS2026submission4281/reviewarena-eval) for the dataset card, evaluation protocol, and baseline results across six open-weight LLMs. ```python from datasets import load_dataset bench = load_dataset("anonymousNeurIPS2026submission4281/reviewarena-eval", split="train") ``` ## Citation Under double-blind review. Please cite the camera-ready NeurIPS proceedings entry once public; until then use an anonymous placeholder consistent with your venue’s reviewer guidelines. ## Terms Use is subject to the redistribution conditions stated on the reviewer-visible Hugging Face dataset card for this submission (including applicable dataset licenses and attribution expectations after deanonymization). ## Repository layout (development snapshot) Anonymized **code** intended for the OpenReview code URL is maintained under **`review-arena/`**. LaTeX sources and submission packaging for the NeurIPS track may appear elsewhere in this repository tree during preparation. ## Acknowledgements - The [OpenReview](https://openreview.net) team for maintaining open peer-review archives. - NVIDIA for [nemotron-ocr-v2](https://huggingface.co/nvidia/nemotron-ocr-v2). - [Modal](https://modal.com) for elastic GPU and storage capacity used in corpus construction.