---
license: cc-by-4.0
language:
- en
pretty_name: researchmap.ai
size_categories:
- 1M<n<10M
task_categories:
- tabular-classification
- text-classification
tags:
- scientometrics
- bibliometrics
- science-of-science
- arxiv
- ai-research
- affiliations
- research-policy
- geography
- citations
configs:
- config_name: papers
  default: true
  data_files:
  - split: train
    path: papers.parquet
- config_name: authorships
  data_files:
  - split: train
    path: authorships.parquet
- config_name: institutions
  data_files:
  - split: train
    path: institutions.parquet
- config_name: references
  data_files:
  - split: train
    path: references.parquet
- config_name: countries
  data_files:
  - split: train
    path: countries.parquet
- config_name: us_university_rd
  data_files:
  - split: train
    path: us_university_rd.parquet
---

# researchmap.ai

Every AI paper posted to arXiv since 2016, with where its authors work. For each of 647,565 papers in cs.AI, cs.LG, cs.CL, cs.CV, cs.NE, cs.RO, cs.MA and stat.ML, the dataset gives its authors' institutions and countries, its subfield, whether it appeared at one of 18 top conferences, and its citations. It is the data behind [researchmap.ai](https://researchmap.ai), which maps where AI research is done.

- **Site and method:** [researchmap.ai](https://researchmap.ai/#method)
- **Covers:** papers through 29 October 2026, citations as of 27 October 2026

![The researchmap.ai map of AI research](assets/site.png)

## Configs

| Config | Rows | One row per |
|---|---:|---|
| `papers` (default) | 647,842 | harvested paper |
| `authorships` | 3,462,906 | paper × author × affiliation |
| `institutions` | 15,590 | institution on at least one listing |
| `references` | 5,505,833 | reference from one paper in the dataset to another |
| `countries` | 2,180 | country × year, 2016–2025 |
| `us_university_rd` | 6,200 | US university × year, 2016–2025 |

The configs join on `paper_id` and `institution_id` (a ROR ID). A null means there is no value. Each table is also a gzipped CSV; 128.0 MB of Parquet in all.

## How it was built

1. **Papers.** Every paper first posted to arXiv since 1 January 2016 in those categories, as primary category or cross-list, from arXiv's OAI-PMH interface.
2. **Affiliations.** Claude Haiku 5.5 read each paper's first page and listed every author's affiliations as printed. An affiliation that doesn't appear in the page text is dropped.
3. **Institutions and countries.** Affiliations are matched to the Research Organization Registry and placed in the country the paper prints, so a lab counts where it is, not where its company has its headquarters.
4. **Credit.** Each paper counts once, split equally among its authors and then between each author's first two affiliations. 5.7% of the credit can't be placed in a country.
5. **Layers.** One of 20 subfields from the title and abstract (the same model); the main track of 18 conferences from official proceedings (DBLP and the conferences' own lists); citations and references from Semantic Scholar, with highly cited meaning the top 10% of a year and subfield; population, GDP and R&D from the World Bank, the OECD and NSF.

## Fields

The main columns. The Hub's viewer shows every column and its type.

| Column | Meaning |
|---|---|
| `papers.year`, `published` | year of the arXiv ID, which the site counts; date version 1 was submitted |
| `papers.subfield`, `subfield_2` | main subfield of 20, and a second one when the paper clearly has one |
| `papers.venue`, `track`, `main_conference` | the top conference and track that published the paper |
| `papers.code_or_project_url`, `code` | the first code or project link on page 1, and whether it is a code host |
| `papers.citations`, `highly_cited` | Semantic Scholar citations; 1 for the top 10% of the year and subfield, a share for ties |
| `papers.countries`, `credit_placed` | the paper's credit by country, and how much of it is placed |
| `authorships.affiliation`, `institution_id` | the affiliation as printed, and the ROR record it matched |
| `authorships.country`, `places` | the country, and where the credit goes (split for a multinational that prints no place) |
| `authorships.sector` | `academia`, `industry` or `independent` |
| `authorships.counted`, `credit` | one of the author's first two affiliations, and its share of the paper |

Countries are ISO codes for where an affiliation is, never anyone's nationality. Hong Kong and Taiwan are kept as papers print them.

## Load

```python
from datasets import load_dataset

papers = load_dataset("ArtKanke/researchmap.ai", "papers", split="train")
authorships = load_dataset("ArtKanke/researchmap.ai", "authorships", split="train")
```

<details>
<summary><b>Example paper</b> (rTaste)</summary>

| Field | Value |
|---|---|
| `paper_id`, `year`, `published` | `2610.40488`, 2026, 2026-10-29 |
| `primary_category`, `n_authors` | `cs.CL`, 1 |
| `subfield`, `subfield_2` | `agents`, `llm_nlp` |
| `code` | true: page 1 links to the code |
| `authorships` | one row: University of Minnesota (`017zqws13`), `US`, `academia`, credit 1.0 |
| `countries`, `credit_placed` | `[{country: US, credit: 1.0}]`, 1.0 |
| `venue`, `citations` | null: too new for a conference list or the monthly citation refresh |

</details>

## License

CC BY 4.0.

## Cite

```bibtex
@misc{kanke2026researchmap,
  author = {Kanke, Art},
  title = {researchmap.ai: arXiv AI papers by institution and country},
  year = {2026},
  url = {https://huggingface.co/datasets/ArtKanke/researchmap.ai}
}
```