PaperLens-Vision

Vision version of the OpenReview-ICLR and arXiv PaperLens datasets.

Each row is one unique paper. We release all extracted papers — not every paper here is used in our downstream training/eval sets. The papers that are used are denoted by the references field, which lists every internal (release_name, release_split) pair the paper belongs to (a single paper can belong to multiple). reconstruction.py reads this field to materialize the original sharegpt data.json for any of the ~20 publishable vision keys.

Configs (subsets)

• arxiv — papers from arxiv (per_venue + 21k families + residual + the arxiv side of combined).
• openreview-iclr — papers from ICLR via OpenReview (balanced_original + max_rejects + train_50pct/75pct + the iclr side of combined).

from datasets import load_dataset
ds_arxiv = load_dataset("skonan/PaperLens-Vision", "arxiv",           split="papers")
ds_iclr  = load_dataset("skonan/PaperLens-Vision", "openreview-iclr", split="papers")

Schema

field	type	description
paper_id	string	arXiv id or OpenReview submission id
title	string	paper title
content	string	prompt-stripped body (title + abstract + <image> placeholders (one per page))
metadata	string	JSON blob — venue, year, authors, ratings, decision, …
label	string	"Accept" or "Reject"
references	list<list<string>>	each entry is [release_name, release_split] — the internal splits this paper belongs to
images	list<struct<bytes,path>>	one PNG per page, bytes inline

Reconstructing the sharegpt data.json files

reconstruction.py rebuilds any of the publishable internal keys (e.g. arxiv_50_50_21k_vision_..._y24up_test) byte-identically from this dataset. Setup + run:

git clone https://github.com/zlab-princeton/PaperLens.git
cd PaperLens/paperlens-training-and-inference
uv sync

# arxiv training set
uv run python scripts/reconstruction.py \
    --hf_vision_repo skonan/PaperLens-Vision \
    --dataset_keys arxiv_50_50_balanced_per_venue_vision_wmetadata_filtered24480_train

# openreview-iclr training set
uv run python scripts/reconstruction.py \
    --hf_vision_repo skonan/PaperLens-Vision \
    --dataset_keys iclr_2020_2023_2025_2026_85_5_10_balanced_original_vision_labelfix_v7_filtered_filtered24480_train

Reconstructed files land in ./data/ by default (override with --data_root <path>): data/<dataset_key>/data.json (sharegpt rows) and data/dataset_info.json (LlamaFactory entry), and data/images_{arxiv,iclr}/<paper_id>/page_*.png for the per-page PNGs.

The release ships a manifest.json sidecar mapping each internal dataset_info.json key → (release_name, release_split, columns, file_name), so reconstruction reproduces conversations, _metadata, accept_reject_label (where applicable), and image bytes exactly.

⚠️ Reconstructing the full vision tree produces one PNG per page, creating over 1M files for the whole release — run on a fileset with inode headroom.

License & citation

License: see the PaperLens collection.

@misc{konan2026paperlens,
  title       = {PaperLens: How Predictable Is Paper Acceptance?},
  author      = {Konan, Sachin and Liu, Jonathan and Liu, Zhuang},
  year        = {2026},
  institution = {Princeton University}
}