debajyotidasgupta/vecforge-paper-corpus

Note (rebuild in progress): figures are being re-extracted with a fixed extractor (cleaner crops). The image-preview config (Parquet with an inline column + difficulty/type/score labels) returns after re-classification. The config (paper metadata + links) is live now.

VecForge Paper Corpus

A pristine, deduplicated collection of 59,732 top-venue AI/ML/CV/NLP/Robotics papers (2020-2024) with every captioned figure and full paper text, for research on figure understanding, editable diagram vectorization, chart/plot analysis, and document parsing. Part of the VecForge project (raster figures -> editable SVG).

Why

Research diagrams - system/architecture overviews, method pipelines, block diagrams, sample galleries, experimental plots - from the best venues are the hard, high-value targets most figure datasets miss.

Sources & coverage (2020-2024)

Field	Venues	Source
CV	CVPR, ICCV, WACV	CVF Open Access
ML	NeurIPS, ICML, AISTATS, CoRL, ICLR	papers.nips.cc, PMLR, OpenReview
Broad	cs.CV/LG/CL/AI/RO, stat.ML	arXiv + HuggingFace daily papers

Deduplicated by arXiv-ID (canonical) then normalized-title hash.

Structure

final_list.json   # deduped master list: {arxiv_id, pdf_url, title, venue, source}
papers.jsonl      # per-paper: {key, title, venue, source, n_figs, pages, text_chars}
figures.jsonl     # per-figure: {id, key, page, caption, wh, venue, source}
figures/          # one pristine PNG per captioned figure ({key}_p{page}_{y}.png)
text/             # full extracted text per paper ({key}.txt)

Figures detected from real graphic content (vector + raster) bounded to each Figure N caption - validated at 100% figure recall with complete, correctly-bounded crops.

Difficulty tiers (curriculum / eval)

• T0 icons/single illustrations (2-5 elt) - T1 pipeline/flow/block (5-15) - T2 dense system/architecture overviews (15-40) - T3 sample galleries, complex multi-panel plots, mixed. (Per-figure tier+type from a VLM pass; see figures.jsonl.)

Usage

from huggingface_hub import hf_hub_download
import json
papers = [json.loads(l) for l in open(hf_hub_download("debajyotidasgupta/vecforge-paper-corpus","papers.jsonl",repo_type="dataset"))]

Licensing & provenance

Research aggregation. Each figure and text retains its original publisher/author license (arXiv per-paper, CVF/IEEE, NeurIPS/PMLR, etc.); shared for non-commercial academic research, not as authorization to redistribute the underlying works. Per-item provenance (source, pdf_url, venue) is kept. Open an issue for takedowns.