README.md

---
license: cc-by-4.0
language:
  - en
task_categories:
  - text-generation
  - summarization
  - feature-extraction
tags:
  - research
  - papers
  - cs
  - machine-learning
  - nlp
  - computer-vision
  - ai
size_categories:
  - 100K<n<1M
configs:
  - config_name: papers
    data_files:
      - split: train
        path: data/papers.jsonl
      - split: arxiv
        path: data/papers_arxiv.jsonl
      - split: conference
        path: data/papers_conference.jsonl
      - split: journal
        path: data/papers_journal.jsonl
  - config_name: instruct
    data_files:
      - split: train
        path: data/instruct.jsonl
  - config_name: sections
    data_files:
      - split: train
        path: data/sections.jsonl
---

# ResearchScope Papers

Open CS research paper dataset maintained by [ResearchScope](https://github.com/kishormorol/ResearchScope).

Updated automatically via GitHub Actions.

## Quick start

```python
from datasets import load_dataset

ds = load_dataset("kishormorol/researchscope-papers", "papers", split="train")
print(ds[0])
```

See [Usage](#usage) below for per-source splits, instruction-tuning, and the per-section fine-tuning data.

## Stats

- **33,079** papers (raw metadata) — **8,079** arXiv · **20,000** conference · **5,000** journal
- **165,181** instruction-tuning rows
- Sources: arXiv, OpenAlex, ACL Anthology, OpenReview, PMLR, CVF, Semantic Scholar
- Venues: NeurIPS, ICML, ICLR, ACL, EMNLP, CVPR, AAAI, IJCAI, JMLR, TMLR, TACL, TPAMI, NMI and more

## Files

| File | Description |
|------|-------------|
| `data/papers.jsonl` | Raw paper metadata — title, abstract, authors, venue, year, tags, scores (all sources combined) |
| `data/papers_arxiv.jsonl` | arXiv / preprint papers only |
| `data/papers_conference.jsonl` | Conference papers only (NeurIPS, ICML, ICLR, ACL, CVPR, …) |
| `data/papers_journal.jsonl` | Journal papers only (JMLR, TPAMI, NMI, TACL, …) |
| `data/instruct.jsonl` | Instruction-tuning pairs — summarize, key contribution, why it matters, plain English |
| `data/sections.jsonl` | Per-section fine-tuning rows for A* papers — real body text of `abstract`, `introduction`, `related_work`, `method`, `experiments`, `results`, `conclusion`. Filter by the `section` field to train a per-section writing agent. |

## Usage

```python
from datasets import load_dataset

# All papers (combined)
papers = load_dataset("kishormorol/researchscope-papers", "papers", split="train")

# Just one source — arXiv, conference, or journal papers
arxiv      = load_dataset("kishormorol/researchscope-papers", "papers", split="arxiv")
conference = load_dataset("kishormorol/researchscope-papers", "papers", split="conference")
journal    = load_dataset("kishormorol/researchscope-papers", "papers", split="journal")

# Instruction tuning
instruct = load_dataset("kishormorol/researchscope-papers", "instruct", split="train")

# Per-section fine-tuning (A* papers) — e.g. train an Introduction-writing agent
sections = load_dataset("kishormorol/researchscope-papers", "sections", split="train")
intros = sections.filter(lambda r: r["section"] == "introduction")
```

## License

Paper metadata is aggregated from open sources. Text content follows the original licenses of each source (arXiv CC0, ACL CC BY, etc.).
Dataset schema: CC BY 4.0.