---
configs:
  - config_name: 1_arxiv_urls_1992_2024_gt
    data_files:
      - split: train
        path: "arxiv_urls_1992_2024_gt.csv"
  # - config_name: 2_arxiv_extracted_unique_urls_all_200_superset
  #   data_files:
  #     - split: train
  #       path: "arxiv_extracted_unique_urls_all_200_superset.json"
  # - config_name: 3_arxiv_extracted_urls_5_formats_360k
  #   data_files:
  #     - split: train
  #       path: "arxiv-mini-corpus/arxiv_extracted_urls_5_formats_360k.json"

license: cc-by-4.0
task_categories:
- text-retrieval
language:
- en
tags:
- arxiv
- scientific-documents
- url-extraction
- pdf-parsing
- reproducibility
- pdf
- latex
- html
- xml
- markdown
- text
- png
pretty_name: arXiv URL Extraction Benchmark & Longitudinal Multi-Format Corpus
size_categories:
- 100K<n<1M
---
# arXiv URL Extraction Benchmark & Longitudinal Corpus

## Dataset Description

This repository hosts datasets designed to assess format-specific coverage gaps in URL extraction across the arXiv corpus. The collection facilitates large-scale reproducibility studies and evaluates how different document formats (LaTeX, HTML, XML, Markdown, TXT, PNG) impact automated extraction pipelines.

The repository is divided into two primary corpora:
1. **The 200-Paper Benchmark:** A curated subset for granular ground-truth analysis.
2. **The 360k Longitudinal Corpus:** A massive, 33-year (1992–2024) parallel corpus where each paper is successfully represented in six distinct formats (PDF, LaTeX, HTML, XML, Markdown, TXT).

### Dataset Sources

### 1. The 200-Paper Benchmark (Ground Truth)
* **Temporal Coverage:** 1992–2024 (Phase A: 2016–24; Phase B: 1992–15)
* **Initial Sampling Pool:** 6,090 PDFs stratified by year and document type
* **Final Verified Pool:** 200 papers containing all 7 formats (`PDF`, `TEXT`, `LaTeX`, `XML`, `HTML`, `Markdown`, `PNG`)
* **Total Extracted Data:** 2,420 total URLs (2,338 unique URLs) manually audited for strict ground truth

### 2. The 360k Longitudinal Corpus
* **Temporal Coverage:** 1992–2024 (Spanning 33 years)
* **Initial Sampling Pool:** 467,467 PDFs (Stratified random sample of up to 15,000 papers per year)
* **Final Common Pool:** 364,744 papers (~7,000 - ~13,000 papers per year) containing all 6 formats (`PDF`, `TEXT`, `LaTeX`, `XML`, `HTML`, `Markdown`)

---

## Dataset Structure & Storage

**Warning:** The full longitudinal corpus (~360k) is exceptionally large. Ensure you have adequate storage and bandwidth before executing a full download.

### 1. The 200-Paper Benchmark Corpus
**File:** [`arxiv-url-bench-200.tar.gz`](https://huggingface.co/datasets/dblind-data/arxiv-url-bench-hf/blob/main/arxiv-url-bench-200.tar.gz)
* **Size:** ~2.8 GB
* **Description:** Contains the raw files for the 200 benchmark papers across their different file format variants.

### 2. The 200-Paper Manual Annotations & OADS Categories
**File:** [`arxiv_urls_1992_2024_gt.csv`](https://huggingface.co/datasets/dblind-data/arxiv-url-bench-hf/blob/main/arxiv_urls_1992_2024_gt.csv)
* **Description:** Contains all manual ground-truth annotations for the 200-paper benchmark, featuring explicit OADS (Open Access Data and Software) category labels along with structural and contextual complexity evaluations for each URL.

### 3. The 200-Paper Unique URLs Superset (including Ground Truth)
**File:** [`arxiv_extracted_unique_urls_all_200_superset.json`](https://huggingface.co/datasets/dblind-data/arxiv-url-bench-hf/blob/main/arxiv_extracted_unique_urls_all_200_superset.json)
* **Size:** ~10 MB
* **Description:** Registry of all extracted URLs across all format variants, including the human-annotated ground truth labels.

<details>
<summary><strong>📋 Show JSON schema for <code>arxiv_extracted_unique_urls_all_200_superset.json</code></strong></summary>

Top-level keys are arXiv IDs. Each entry contains one sub-object per extraction source (formats variants, VLM outputs, and the merged superset):

```json
{
  "arxiv_id": {
    "text": {
      "url_count": "int",
      "filename": "str",
      "urls": ["str"]
    },
    "html": {
      "url_count": "int",
      "filename": "str",
      "urls": ["str"]
    },
    "latex": {
      "url_count": "int",
      "urls": ["str"]
    },
    "xml": {
      "url_count": "int",
      "filename": "str",
      "urls": ["str"]
    },
    "vlm_qwen": {
      "url_count": "int",
      "filename": "str",
      "urls": ["str"]
    },
    "vlm_deepseek": {
      "url_count": "int",
      "filename": "str",
      "urls": ["str"]
    },
    "vlm_minicpm": {
      "url_count": "int",
      "filename": "str",
      "urls": ["str"]
    },
    "textwal": {
      "url_count": "int",
      "filename": "str",
      "urls": ["str"]
    },
    "textwal_pypdf": {
      "url_count": "int",
      "filename": "str",
      "urls": ["str"]
    },
    "textwal_pdfminer": {
      "url_count": "int",
      "filename": "str",
      "urls": ["str"]
    },
    "textwalcl": {
      "urls": ["str"],
      "number_of_pages": "int",
      "input_tokens": "int",
      "output_tokens": "int",
      "cost_usd": "float",
      "runtime_sec": "float",
      "failed": "bool"
    },
    "markdown": {
      "filename": "str",
      "md_path": "str",
      "url_count": "int",
      "urls": ["str"]
    },
    "pdf": {
      "urls": ["str"],
      "url_count": "int"
    },
    "oads": {
      "oads_info": ["str"],
      "url_count": "int"
    },
    "superset": {
      "urls": ["str"],
      "url_count": "int"
    }
  }
}
```

**Field notes:**
* `text`, `html`, `xml`, `latex`, `markdown`, `pdf` — per-format extraction results, each with the URLs found (`urls`), the count (`url_count`), and, where applicable, the source `filename`/`md_path`.
* `vlm_qwen`, `vlm_deepseek`, `vlm_minicpm` — extraction results from the respective vision-language models run over the paper's page images.
* `textwal`, `textwal_pypdf`, `textwal_pdfminer` — text-based extraction variants using different underlying parsing libraries.
* `textwalcl` — Claude-based extraction (token usage, cost, runtime, failure status) in addition to the URLs found.
* `oads` — Open Access Data and Software category info for the URLs in this entry.
* `superset` — the deduplicated union of URLs across all extraction sources for this paper.

</details>

### 4. The 360k Longitudinal Multi-Format Corpus
**Files:** [`pdf.tar.gz,latex.tar.gz,html.tar.gz,textwal.tar.gz,xml.tar.gz,markdown.tar.gz,`](https://huggingface.co/datasets/dblind-data/arxiv-url-bench-hf/tree/main/arxiv-mini-corpus)
* **Size:** ~560 GB
* **Description:** The complete, multi-format synchronized common pool containing 364,744 papers.

### 5. The 360k Longitudinal Extracted URLs
**File:** [`arxiv_extracted_urls_5_formats_360k.json`](https://huggingface.co/datasets/dblind-data/arxiv-url-bench-hf/blob/main/arxiv-mini-corpus/arxiv_extracted_urls_5_formats_360k.json)
* **Size:** ~690 MB
* **Description:** Compiled output of every URL extracted from the 360k dataset across the 5 core text formats.

<details>
<summary><strong>📋 Show JSON schema for <code>arxiv_extracted_urls_5_formats_360k.json</code></strong></summary>

Top-level keys are publication years, each mapping to arXiv IDs published that year. Each arXiv ID contains one sub-object per format, with the extracted URLs and a count:

```json
{
  "year": {
    "arxiv_id": {
      "textwal": {
        "urls": ["str"],
        "num_urls": "int"
      },
      "latex": {
        "urls": ["str"],
        "num_urls": "int"
      },
      "xml": {
        "urls": ["str"],
        "num_urls": "int"
      },
      "markdown": {
        "urls": ["str"],
        "num_urls": "int"
      },
      "html": {
        "urls": ["str"],
        "num_urls": "int"
      }
    }
  }
}
```

**Field notes:**
* `year` — four-digit publication year (e.g. `"2016"`), used to shard the 364,744 papers.
* `arxiv_id` — arXiv identifier of the paper.
* `textwal`, `latex`, `xml`, `markdown`, `html` — one sub-object per format, each holding the list of extracted `urls` and the corresponding `num_urls` count.

</details>

---