README.md

---
pretty_name: Scientific Contribution Graph (SCG)
language:
  - en
license: cc-by-4.0
task_categories:
  - graph-ml
  - text-generation
  - other
task_ids:
  - entity-linking-classification
tags:
  - scientific-literature
  - knowledge-graph
  - technological-roadmapping
  - prerequisite-prediction
  - scholarly-documents
  - nlp
  - science
  - s2orc
  - acl-anthology
  - parquet
size_categories:
  - 1M<n<10M
source_datasets:
  - acl-anthology
  - s2orc
multilinguality:
  - monolingual
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/**/*.parquet
---

# Dataset Card for the Scientific Contribution Graph (SCG)

> **Compressed Parquet release:** this Hub dataset is a compact Parquet encoding of the Scientific Contribution Graph.  
> It reduces the on-disk footprint from **~80 GB uncompressed JSON/tree files → ~6 GB Parquet**, while preserving the same contributions, prerequisite edges, and metadata as SCG **v1.1**.

The **Scientific Contribution Graph** maps how science is built “on the shoulders of giants”: it extracts fine-grained scientific contributions from open-access papers and links each contribution to the precursor contributions it builds on.

This card documents the **v1.1** graph in Parquet form. The original full-tree release used by the [SCG Explorer](https://huggingface.co/spaces/pajansen/scientific-contribution-graph) and the official API lives in the bucket [`pajansen/scg-explorer-space-data`](https://huggingface.co/buckets/pajansen/scg-explorer-space-data) (~84.4 GB, 655,252 files; ~80 GB uncompressed / ~19 GB tar-compressed). **Prefer this Parquet dataset when you want the same graph at ~6 GB.**

Paper: [The Scientific Contribution Graph: Automated Literature-based Technological Roadmapping at Scale](https://arxiv.org/abs/2605.15011) (Findings of EMNLP 2026).

## Dataset Details

### Dataset Description

Scientific contributions rarely appear in isolation. SCG treats automated **technological roadmapping** as knowledge-graph construction over full paper text:

- **Nodes** are individual scientific contributions (name + long-form description + metadata), not whole papers and not short entity spans.
- **Edges** are prerequisite / precursor relations, each with a natural-language explanation of *why* one contribution enables another, plus alignment to cited papers.

Extraction is modeled as open-ended sequence-to-sequence generation over full text (on the order of ~43 LLM calls per paper), then cross-paper alignment of prerequisites to contributions in cited work.

| Representation | Typical size | Notes |
| --- | ---: | --- |
| Original v1.1 tree / JSON | **~80 GB** uncompressed (~19 GB `.tar.gz`) | Per-paper files in [`pajansen/scg-explorer-space-data`](https://huggingface.co/buckets/pajansen/scg-explorer-space-data); native API format |
| **This dataset (Parquet)** | **~6 GB** | Columnar, compressed encoding of the same v1.1 graph |

| Release | Papers | Contributions | Prerequisite edges | Domain |
| --- | ---: | ---: | ---: | --- |
| v1.0 (May 2026) | 230,454 | 2,047,426 | 12,524,458 | ACL Anthology seed + S2ORC back-crawl (largely AI/NLP) |
| **v1.1 (Aug 2026, current)** | **655,195** | **5,976,913** | **36,199,762** | Broader science via alternating forward/backward crawls |

Averages (v1.1):

- ~9 contributions per paper
- ~6 prerequisites per contribution
- ~130 tokens per contribution description
- ~71 tokens per edge explanation
- Prerequisite mix: ~61% external papers, ~34% intra-paper contributions, ~5% tools/URLs
- Linked paper-reference → specific downstream contribution: 29.0% (up from 25.7% in v1.0)

**Curated by:** Peter A. Jansen (University of Arizona; Allen Institute for Artificial Intelligence)  
**Language(s):** English (source papers are English scholarly text)  
**License:** [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/) for the extracted graph annotations and paper. Underlying open-access papers retain their original licenses.

### Dataset Sources

- **Paper:** https://arxiv.org/abs/2605.15011  
- **DOI:** https://doi.org/10.48550/arXiv.2605.15011  
- **Code / API:** https://github.com/cognitiveailab/scientific-contribution-graph  
- **Live demo:** https://huggingface.co/spaces/pajansen/scientific-contribution-graph  
- **Original (uncompressed) data bucket:** https://huggingface.co/buckets/pajansen/scg-explorer-space-data  
- **Source corpora:** [ACL Anthology](https://aclanthology.org/) (Feb 2025 snapshot for v1.0 seed) and [Semantic Scholar Open Research Corpus (S2ORC)](https://www.semanticscholar.org/)

## Uses

### Direct Use

- Loading the full SCG on a laptop or modest server without provisioning ~80 GB of JSON
- Technological roadmapping and contribution-level citation / dependency analysis
- Scientific impact assessment at *contribution* granularity (not only paper-level citations)
- Training or evaluating **scientific prerequisite prediction** (which existing technologies enable a future contribution)
- Graph mining, visualization, and retrieval over scientific progress
- Input for automated scientific discovery systems that reason about enabling technologies

### Out-of-Scope Use

- Treating extracted nodes/edges as gold human annotations or as a complete map of a field (coverage is limited to open-access full text that the crawl reached)
- Using prerequisite edges as legal claims of priority, inventorship, or plagiarism
- Assuming every citation in a paper is a true technological prerequisite (the graph is designed to be finer-grained than raw citations, but alignment is imperfect)
- Deploying rankings or “impact scores” as sole criteria for hiring, funding, or clinical/scientific decisions without human review

## Dataset Structure

This repository stores SCG v1.1 as **Parquet** rather than one JSON document per paper. Content is equivalent to the official release; only the on-disk encoding changes (~80 GB → ~6 GB).

### Identifiers

- Papers are keyed by **Semantic Scholar corpus IDs**.
- Contributions are `"{corpus_id}.c{n}"` (e.g. `52967399.c1`).

### Conceptual schema

Each **contribution node** is a tuple \(\langle n_i, d_i, m_i \rangle\):

- `name` — short contribution title  
- `description` — detailed generated summary  
- `metadata` — source paper title, sections mentioning the contribution, etc.

Each **prerequisite edge** is \(\langle c_{\mathrm{pre}}, c_{\mathrm{dep}}, a_j \rangle\):

- precursor and dependent contribution IDs  
- attributes such as a natural-language justification and the cited paper used for alignment

The official Python API in the GitHub repo still expects the original per-paper tree plus `data/metadata/paper_manifest.jsonl`. Use this Parquet dump for analysis with `datasets`, DuckDB, Polars, or Spark; convert back to the tree layout only if you need the demo API unchanged.

### Topic distribution (v1.1)

Semantic Scholar topic labels (papers may have multiple topics; percentages sum to >100%):

| Topic | Count | % | Topic | Count | % |
| --- | ---: | ---: | --- | ---: | ---: |
| Medicine | 304,898 | 46.5 | Sociology | 4,796 | 0.7 |
| Computer Science | 272,796 | 41.6 | Education | 3,637 | 0.6 |
| Biology | 177,090 | 27.0 | Political Science | 3,108 | 0.5 |
| Physics | 163,066 | 24.9 | Business | 2,642 | 0.4 |
| Mathematics | 58,616 | 8.9 | Economics | 2,628 | 0.4 |
| Linguistics | 51,830 | 7.9 | Agricultural and Food Sciences | 2,615 | 0.4 |
| Engineering | 51,458 | 7.9 | Geography | 1,904 | 0.3 |
| Environmental Science | 40,636 | 6.2 | Philosophy | 1,900 | 0.3 |
| Chemistry | 35,962 | 5.5 | Law | 1,650 | 0.3 |
| Materials Science | 20,198 | 3.1 | History | 1,111 | 0.2 |
| Psychology | 13,494 | 2.1 |  |  |  |

Topics with &lt;1,000 papers omitted (e.g. Art, Geology).

### Data Splits

The graph release itself is a single corpus (no official train/dev/test split of nodes/edges).

The companion **scientific prerequisite prediction** task uses **temporally filtered backtesting** (predict precursors of later contributions from earlier graph state). Reported headline result in the paper: **0.48 MAP**. Task code/data: `task_precursor_prediction/` in the GitHub repository.

## Dataset Creation

### Curation Rationale

Paper-level citations mix background mentions with true “built upon” links. Sentence-level IE triples are usually too short and decontextualized. SCG sits in between: contribution-level nodes with long descriptions, and explained prerequisite edges.

The Parquet packaging exists so the same 6M-node / 36M-edge graph can be distributed and scanned without an 80 GB checkout.

### Source Data

**v1.0:** full ACL Anthology crawl, then backward crawl through S2ORC along cited open-access papers.

**v1.1:** alternating **forward** (impact) and **backward** (dependency) crawls through S2ORC; reference linking improved by backfilling missing S2 corpus IDs via exact title match, then re-running cross-contribution alignment.

#### Who produced the source language?

Authors of the underlying scholarly articles. Graph text (names, descriptions, edge explanations) is **model-generated** from those papers’ full text.

### Annotations

Automatically generated; not crowd-annotated. Human evaluation and alignment statistics are in the paper. Residual errors include missed contributions, spurious prerequisites, and failed or coarse paper–contribution alignment (~71% of paper references still not pinned to a specific downstream contribution).

### Personal and Sensitive Information

Records are scientific contributions from published open-access papers (titles, IDs, generated technical summaries). Author names may appear as in the source literature. No intent to include private personal data; biomedical papers may discuss clinical topics at the level already present in the publications.

## Considerations for Using the Data

### Social Impact

Useful for literature review, impact analysis, and discovery systems. Risk of amplifying venue/OA biases or treating generated “roadmaps” as authoritative history of science.

### Discussion of Biases

- Open-access and crawl-seed bias (ACL/NLP origin, then expansion)
- English-language scholarly bias
- Field mix in v1.1 is heavy in medicine