codeofpaper/paper-code-graph

1---2license: cc-by-4.03language:4  - en5pretty_name: codeofpaper — paper ↔ code graph6size_categories:7  - 100K<n<1M8task_categories:9  - text-retrieval10  - feature-extraction11tags:12  - reproducibility13  - github14  - arxiv15  - research-software16  - paper-code-linking17  - scientific-papers18  - source-code19  - knowledge-graph20  - benchmark21configs:22  - config_name: papers23    data_files: papers.parquet24  - config_name: repositories25    data_files: repositories.parquet26  - config_name: paper_repositories27    data_files: paper_repositories.parquet28  - config_name: repository_forks29    data_files: repository_forks.parquet30---31 32# codeofpaper: paper ↔ code graph33 34A high-precision dataset linking research papers (arXiv, OpenReview) to their35**official** and **high-confidence community** GitHub implementations, plus36the full fork lineage for those repositories. Built and maintained by the37[codeofpaper](https://codeofpaper.com) project.38 39**Snapshot date:** 2026-05-1940**Schema version:** 1.0.041**License:** CC-BY-4.0 (metadata aggregation + derived tier labels).42Underlying paper texts and repository code remain under their own licenses;43this dataset only ships identifiers, URLs, and structured metadata.44 45## What's in here46 47- **Tiered repo links**: every edge is classified as `official` or48  `high_confidence_community` by a reproducible classifier. Lower-confidence49  edges (`possible_match`) are intentionally excluded from the published50  snapshot to keep precision high.51- **Fork lineage**: `repository_forks` covers ~1.5M fork edges across ~89k52  parent repos, enabling research-fork discovery: hardware ports, quantization53  variants, lab extensions, and other derivative work that's normally54  invisible in paper ↔ code link tables.55- **Engineering metadata**: `repositories.framework`, `license_spdx`,56  `engineering_fields_evidence` (JSONB) carry the signals our tier classifier57  consumes, so you can audit or train your own.58 59## Files60 61| File | Rows | Description |62|------|------|-------------|63| `papers.parquet` | 244,180 | Paper metadata (arXiv / OpenReview / DOI). |64| `repositories.parquet` | 59,296 | Non-blocklisted GitHub repos linked to at least one paper. |65| `paper_repositories.parquet` | 62,634 | Confident-tier edges only (`official` ∪ `high_confidence_community`). |66| `repository_forks.parquet` | 543,340 | Fork lineage for parents in `repositories.parquet`. |67 68## Schema notes69 70- All `*_evidence`, `languages`, and other JSONB columns are serialized as71  compact JSON strings (parquet doesn't natively map nested types well across72  loaders). Parse with `json.loads()` in Python.73- `paper_repositories.tier_evidence` documents *why* the classifier assigned a74  tier, useful for auditing or retraining your own classifier.75- A 90-day cutoff on `papers.published_date` is applied. The most recent76  papers (and their newly-linked repos) intentionally aren't in this snapshot.77 78## Quickstart79 80```python81from datasets import load_dataset82 83papers = load_dataset("codeofpaper/paper-code-graph", name="papers", split="train")84repos = load_dataset("codeofpaper/paper-code-graph", name="repositories", split="train")85links = load_dataset("codeofpaper/paper-code-graph", name="paper_repositories", split="train")86forks = load_dataset("codeofpaper/paper-code-graph", name="repository_forks", split="train")87 88# Find the official implementation for a paper89import polars as pl90df_links = pl.from_arrow(links.data.table)91df_repos = pl.from_arrow(repos.data.table)92(df_links.filter(pl.col("tier") == "official")93         .join(df_repos, left_on="repository_id", right_on="id")94         .select(["paper_id", "full_name", "stars", "framework"]))95```96 97## Refresh cadence98 99Republished monthly. Production data updates daily on100[codeofpaper.com](https://codeofpaper.com), but the public snapshot ships101on a 90-day delay to avoid freshness contracts.102 103## Citation104 105If you use this dataset, please cite:106 107```bibtex108@misc{codeofpaper_2026,109  title = {codeofpaper: paper-code graph},110  author = {codeofpaper contributors},111  year = {2026},112  url = {https://huggingface.co/datasets/codeofpaper/paper-code-graph},113}114```115 116## Project117 118Product: <https://codeofpaper.com>. The classifier rubric, build script, and119methodology will be open-sourced alongside the first public release of the120codeofpaper repository.121 122Build metadata: schema `1.0.0`, git SHA `b7ae3a7eace01038093aff1918b9c09754dd088a`.123