codeofpaper/paper-code-graph

codeofpaper: paper ↔ code graph

A high-precision dataset linking research papers (arXiv, OpenReview) to their official and high-confidence community GitHub implementations, plus the full fork lineage for those repositories. Built and maintained by the codeofpaper project.

Snapshot date: 2026-05-19 Schema version: 1.0.0 License: CC-BY-4.0 (metadata aggregation + derived tier labels). Underlying paper texts and repository code remain under their own licenses; this dataset only ships identifiers, URLs, and structured metadata.

What's in here

• Tiered repo links: every edge is classified as official or high_confidence_community by a reproducible classifier. Lower-confidence edges (possible_match) are intentionally excluded from the published snapshot to keep precision high.
• Fork lineage: repository_forks covers ~1.5M fork edges across ~89k parent repos, enabling research-fork discovery: hardware ports, quantization variants, lab extensions, and other derivative work that's normally invisible in paper ↔ code link tables.
• Engineering metadata: repositories.framework, license_spdx, engineering_fields_evidence (JSONB) carry the signals our tier classifier consumes, so you can audit or train your own.

Files

File	Rows	Description
papers.parquet	244,180	Paper metadata (arXiv / OpenReview / DOI).
repositories.parquet	59,296	Non-blocklisted GitHub repos linked to at least one paper.
paper_repositories.parquet	62,634	Confident-tier edges only (official ∪ high_confidence_community).
repository_forks.parquet	543,340	Fork lineage for parents in repositories.parquet.

Schema notes

• All *_evidence, languages, and other JSONB columns are serialized as compact JSON strings (parquet doesn't natively map nested types well across loaders). Parse with json.loads() in Python.
• paper_repositories.tier_evidence documents why the classifier assigned a tier, useful for auditing or retraining your own classifier.
• A 90-day cutoff on papers.published_date is applied. The most recent papers (and their newly-linked repos) intentionally aren't in this snapshot.

Quickstart

from datasets import load_dataset

papers = load_dataset("codeofpaper/paper-code-graph", name="papers", split="train")
repos = load_dataset("codeofpaper/paper-code-graph", name="repositories", split="train")
links = load_dataset("codeofpaper/paper-code-graph", name="paper_repositories", split="train")
forks = load_dataset("codeofpaper/paper-code-graph", name="repository_forks", split="train")

# Find the official implementation for a paper
import polars as pl
df_links = pl.from_arrow(links.data.table)
df_repos = pl.from_arrow(repos.data.table)
(df_links.filter(pl.col("tier") == "official")
         .join(df_repos, left_on="repository_id", right_on="id")
         .select(["paper_id", "full_name", "stars", "framework"]))

Refresh cadence

Republished monthly. Production data updates daily on codeofpaper.com, but the public snapshot ships on a 90-day delay to avoid freshness contracts.

Citation

If you use this dataset, please cite:

@misc{codeofpaper_2026,
  title = {codeofpaper: paper-code graph},
  author = {codeofpaper contributors},
  year = {2026},
  url = {https://huggingface.co/datasets/codeofpaper/paper-code-graph},
}

Project

Product: https://codeofpaper.com. The classifier rubric, build script, and methodology will be open-sourced alongside the first public release of the codeofpaper repository.

Build metadata: schema 1.0.0, git SHA b7ae3a7eace01038093aff1918b9c09754dd088a.