File size: 5,026 Bytes
78bbdba | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 | ---
pretty_name: SurveyReview
language:
- en
tags:
- survey-evaluation
- peer-review
- benchmark
- text
task_categories:
- text-classification
- text-generation
---
# SurveyReview
SurveyReview is a reviewer-aligned benchmark for evaluating survey papers. It turns real peer-review reports into multidimensional scores and rationales so that model judgments can be compared with human reviewer judgments.
The benchmark covers four dimensions: **Readability**, **Criticalness**, **Comprehensiveness**, and **Structure**.
> **Latest release: [`v1.1`](https://huggingface.co/datasets/brighterrluo/SurveyReview/tree/main/v1.1)**
## What's New in v1.1
- Cleaned full-text content for 1,646 survey articles, stored in two JSON shards.
- The original train/test labels and benchmark dimensions are preserved.
- Evaluation prompts now request structured `<reason>...</reason> <score>X</score>` responses.
- An API-based evaluation pipeline reports MSE and MAE for scores and can optionally use a judge model to compute Rationale Quality Score (RQS).
- Lightweight LLaMA-Factory recipes are included for Qwen3-32B LoRA supervised fine-tuning on each of the four dimensions.
See the [v1.1 release README](https://huggingface.co/datasets/brighterrluo/SurveyReview/blob/main/v1.1/README.md) for setup, evaluation, and benchmark results.
## Versions
| Version | Status | Description |
| --- | --- | --- |
| [`v1.1/`](https://huggingface.co/datasets/brighterrluo/SurveyReview/tree/main/v1.1) | Latest | Cleaned article texts, XML-style rationale/score prompts, evaluation code, and training recipes. |
| [`v1.0-paper/`](https://huggingface.co/datasets/brighterrluo/SurveyReview/tree/main/v1.0-paper) | Archived | Paper-aligned data plus Marker Markdown and Grobid XML resources for reproduction. |
## v1.1 Layout
| Path | Contents |
| --- | --- |
| `v1.1/data/v1.1-paper/raw/` | Review-level train and test samples used for the reported data statistics. |
| `v1.1/data/v1.1-paper/train/` | Training labels grouped by survey. |
| `v1.1/data/v1.1-paper/test/` | Test labels grouped by survey. |
| `v1.1/data/v1.1-paper/articles/` | Cleaned article full texts, split into two JSON shards. |
| `v1.1/data/v1.1-paper/prompt/` | Dimension definitions, evaluation prompts, and the RQS judge prompt. |
| `v1.1/src/` | OpenAI-compatible API evaluation pipeline. |
| `v1.1/training/` | Data conversion script and LLaMA-Factory Qwen3-32B LoRA configurations. |
## Data Statistics
| Item | Count |
| --- | ---: |
| Raw train review samples | 1,216 |
| Raw test review samples | 414 |
| Grouped train surveys | 480 |
| Grouped test surveys | 163 |
| Cleaned article full texts | 1,646 |
Each grouped record contains a survey identifier, title and abstract, review text, source, multilingual metadata where available, and a list of dimension-level scores and reviewer rationales. The article shards map each survey identifier to its cleaned full text.
The evaluator uses scores in `{-2, -1, 1, 2}`. Unlabeled or non-evaluated entries with score `0` or `-3` are skipped by the provided evaluation and training utilities.
## Download v1.1
```python
from huggingface_hub import snapshot_download
snapshot_download(
repo_id="brighterrluo/SurveyReview",
repo_type="dataset",
allow_patterns=["v1.1/**"],
local_dir="SurveyReview",
)
```
## Run the v1.1 Evaluator
```bash
cd SurveyReview/v1.1
python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
cp .env.example .env
```
Set `API_KEY`, `BASE_URL`, `MODEL_NAME`, and `JUDGE_MODEL` in `.env`, then run:
```bash
python src/api_base_evaluate.py
```
The evaluator uses the test split by default. Set `EVAL_SPLIT=train` to evaluate the training split, or set `EVALUATE_REASONS=False` to skip judge-model rationale scoring.
## v1.0-paper Resources
The archived `v1.0-paper` release includes the original review data, prompts, article data, 1,646 Marker-parsed Markdown articles, and 1,646 Grobid XML articles.
To reconstruct and extract the Marker Markdown archive:
```bash
cat v1.0-paper/markdown_archive/marker_markdown.tar.gz.part-* > marker_markdown.tar.gz
tar -xzf marker_markdown.tar.gz
```
To extract the Grobid XML archive:
```bash
tar -xzf v1.0-paper/grobid_xml_archive/grobid_xml.tar.gz
```
## Citation
If you use SurveyReview in your research, please cite:
```bibtex
@inproceedings{zhang2026surveyreview,
author = {Zhang, Yuheng and Wang, Yuanchun and Zhang, Fanjin and Zhao, Ruyu and Li, Juanzi and Tang, Jie and Zhang, Jing},
title = {{SurveyReview}: A Reviewer-Aligned Benchmark for Survey Evaluators},
booktitle = {Proceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining V.2},
year = {2026},
pages = {10302--10313},
publisher = {Association for Computing Machinery},
address = {New York, NY, USA},
location = {Jeju Island, Republic of Korea},
series = {KDD '26},
isbn = {979-8-4007-2259-2},
doi = {10.1145/3770855.3817505},
url = {https://doi.org/10.1145/3770855.3817505}
}
```
|