NovGauge

NovGauge evaluates paper similarity along three dimensions: task, problem, and method. This dataset contains final benchmark labels and bibliographic metadata for pairwise classification and multi-paper grouping.

Dataset repository: ZitaGo/NovGauge.

Release contents

File	Records	Contents
data/positives.json	463	Paper pairs similar in at least one annotated dimension
data/negatives.json	156	Paper pairs dissimilar in the specified annotated dimensions
data/grouping.json	50	Multi-paper sets with dimension-specific reference groups

There are 619 paper pairs, including 236 ICLR-derived positive pairs, 227 survey-derived positive pairs and 156 survey-derived negative pairs. All samples belong to the test split. No training or development split is created in this release.

Of the 50 grouping sets, 44 have at least one evaluated dimension; 6 have an empty eval_dims list and contribute no scored item-dimension tasks.

Only identifiers, bibliographic metadata and final labels are distributed. Paper PDFs, abstracts, introductions, full text, evidence quotations, review text, annotation rationales, local paths, parsing caches, internal service links and construction scripts are excluded. The model evaluation framework is a separate release; its repository link will be added when finalized.

Load with Hugging Face Datasets

After the dataset files are published, load them with:

from datasets import load_dataset

dataset_id = "ZitaGo/NovGauge"
pairwise = load_dataset(dataset_id, "pairwise", split="test")
grouping = load_dataset(dataset_id, "grouping", split="test")

The pairwise configuration concatenates the positive and negative files in that order. Each JSON file is also an ordinary UTF-8 JSON array that can be read with json.load. No dataset loading script or remote code is required. For reproducible comparisons, pin the published dataset commit using the revision argument.

Pairwise schema

Field	Meaning
pair_id	Unique identifier for a paper-pair sample
source	iclr or survey
paper_a, paper_b	Bibliographic paper records described below
labels	Object with task, problem, and method; each value is 1, 0, or null
source_survey	Source survey arXiv identifier, or null for ICLR-derived pairs

Label semantics:

• 1: similar in that dimension.
• 0: dissimilar in that dimension.
• null: unannotated / unevaluated. It is not a negative label.

For positive pairs, only the dimensions marked 1 in labels are positive. Other dimensions remain null. For negative pairs, only the explicitly annotated dimensions are zero. In particular, a pair in negatives.json is not automatically negative in all three dimensions.

Dimension	Positive labels	Negative labels	Total scored labels
Task	222	77	299
Problem	162	33	195
Method	265	116	381
Total	649	226	875

These are dimension-label counts, not counts of distinct pairs. One pair can have more than one annotated dimension.

All 619 pair_id values are unique. Use each complete identifier for evaluation and resume keys; do not truncate it or remove suffixes.

Grouping schema

Field	Meaning
item_id	Grouping sample identifier derived from its source annotation
source_survey	ArXiv identifier of the source survey
n_papers	Number of papers in this set
papers	Ordered paper records, additionally containing paper_index
ground_truth	task_groups, problem_groups, and method_groups, each a list of index lists
eval_dims	The dimensions that are valid for scoring this set

paper_index is one-based and refers to the order in papers. Preserve this order, the reference groups and eval_dims when preparing evaluation inputs. Do not infer a new negative label from an unevaluated dimension.

The numbers of valid item-dimension tasks are 44 for task, 13 for problem and 28 for method, for a total of 85.

Bibliographic paper records

Field	Meaning
paper_id	Stable identifier within this release, formatted as ngp-000001
title	Bibliographic title
authors	Bibliographic author strings; an empty list means unavailable
year, venue	Available publication metadata, otherwise null
arxiv_id	Normalized arXiv identifier, retaining an explicit version suffix when supplied
doi, forum_id, s2_paper_id	Available external identifiers, otherwise null
source_url	Public bibliographic or paper URL, otherwise null

There are 692 benchmark paper records. paper_id identifies a record used by the benchmark, not a guaranteed unique publication: different records may refer to the same work or different representations of it. Records have not been merged merely because their titles or external identifiers match.

18 paper records have no available public source_url; their other available metadata remain in the release. Source links are not guaranteed to remain live. An unversioned identifier does not specify an exact text version. Acquire content matching the bibliographic records and record the versions used for evaluation.

Use with the evaluation framework

Download the three JSON files and retain their labels, sample IDs and paper ordering. Locally add a markdown_path field to each paper record, pointing to your own corresponding Markdown file. For abstract-based modes, also supply the local abstract field. These content fields are deliberately not distributed by this dataset.

The files use the evaluation framework's paper_a / paper_b, labels, and grouping schemas. labels is the single pairwise label representation. Point the pairwise workflow to the positive and negative files, and the grouping workflow to grouping.json. Paper acquisition and PDF-to-Markdown conversion are the user's responsibility. The internal parsing API used in the original experiments is not included; different paper versions or parsing outputs can change model inputs and scores.

Provenance and intended use

ICLR-derived positive pairs originate from paper-overlap annotations associated with OpenReview material. Survey-derived positives, negatives and grouping sets come from curated survey annotations. Survey negatives follow the dimension-specific cross-group rule.

The intended use is evaluation of dimension-specific similarity and grouping. An unannotated dimension is not evidence of novelty, and these labels do not provide a universal novelty ranking or a publication-acceptance decision. The benchmark is concentrated in the scientific domains and papers represented by its sources; metadata and labels can contain errors. Public test labels can also become training contamination, so their use in training should be disclosed when reporting evaluation results.

License

This dataset is released under Creative Commons Attribution 4.0 International (CC BY 4.0). The license applies to the dataset release; third-party paper content retains its respective licenses.