k-chirkunov/nlpeer_v2_verification_dataset_qwen36_35b_a3b

NLPeer v2 — review-weakness verification dataset

One row per weakness point raised in a peer review, paired with two versions of the reviewed paper: the full paper, and the same paper with the authors' claims/evidence for that specific weakness point removed. A model asked to verify the point against the second version can no longer find the text that settles it, which makes the pair usable for testing whether a verifier actually grounds its judgement in the paper.

Built from NLPeer v2 by the pipeline in nlpeer_v2_verification_dataset_generation/ (see build_dataset.py / pipeline.py); the evidence removal is done with qwen3.6-35b-a3b, hence the repository name.

Fields

field	type	description
nlpeer_v2_paper_id	string	NLPeer v2 paper uid (the top-level folder name in the source dataset)
nlpeer_v2_review_id	string	review id (rid in the paper's reviews.json) the weakness point comes from
weakness_point_id	string	wp-<n>, where n indexes the point over all non-empty lines of the review's summary_of_weaknesses. Numbering has gaps: points dropped by the pipeline (not self-contained, not checkable, or no evidence found) get no row, and the id always maps back to the same line of the original review
paper_verifiable_itg_json	string	the unmodified paper as an Intertext Graph document, serialised JSON ({"nodes": [...], "edges": [...]}). "Verifiable" = the evidence for this weakness point is still present
paper_non_verifiable_itg_json	string	the same ITG document with the nodes carrying the authors' claims/evidence for this weakness point removed. "Non-verifiable" = the point can no longer be settled from the paper
weakness_point	string	the weakness-point text, verbatim from the review

Both ITG fields are stored as raw JSON strings (not nested structs), so parse them with json.loads — the node schemas vary per paper and would not survive a fixed Arrow schema. Every field is non-null and non-empty in every row.

The two ITG documents differ only in the removed nodes; paper_verifiable_itg_json repeats per row of the same paper, so filter to unique nlpeer_v2_paper_id if you only need the papers.

Size

• 773 rows — 364 papers, 439 reviews
• rows per paper: one per weakness point that survived the pipeline (papers whose reviews yielded no verifiable weakness point are absent)

Usage

import json
from datasets import load_dataset

ds = load_dataset("k-chirkunov/nlpeer_v2_verification_dataset_qwen36_35b_a3b", split="train")

row = ds[0]
full = json.loads(row["paper_verifiable_itg_json"])
ablated = json.loads(row["paper_non_verifiable_itg_json"])
print(row["weakness_point"])
print(len(full["nodes"]), "->", len(ablated["nodes"]), "nodes")

Provenance and terms

Papers and reviews come from NLPeer v2; their original licences and terms of use apply to the content reproduced here. The per-weakness-point ablations are model-generated and were not manually verified — treat removals as best-effort rather than exhaustive.