README.md

---
language:
  - es
license: cc-by-4.0
tags:
  - history
  - spain
  - 23-F
  - declassified-documents
  - transcripts
  - coup-attempt
  - education
  - spanish-transition
  - document-ai
  - ocr
task_categories:
  - image-to-text
  - text-generation
  - question-answering
pretty_name: "Expediente 23-F: Declassified Documents from the 1981 Spanish Coup Attempt"
size_categories:
  - 1K<n<10K
---

# Expediente 23-F: Declassified Documents from the 1981 Spanish Coup Attempt

A page-level multimodal dataset of declassified Spanish government documents related to the failed coup d'etat of February 23, 1981 (known as 23-F). Each row contains a rendered page image paired with its extracted text, sourced from official state archives.

## Historical Context

On February 23, 1981, Lieutenant Colonel Antonio Tejero stormed the Spanish Congress of Deputies with armed Guardia Civil officers during the investiture vote of Prime Minister Leopoldo Calvo-Sotelo. The coup failed after King Juan Carlos I publicly opposed it in a televised address. These documents were declassified and published by the Spanish Government through the *Portal de Archivos Estatales*.

## Dataset Schema

| Column | Type | Description |
|---|---|---|
| `image` | `Image` | Rendered page image (150 DPI) |
| `pdf_source` | `string` | Relative path to the original PDF or image file |
| `page_number` | `int32` | Page number within the source document (1-indexed) |
| `text` | `string` | Extracted text for the page (OCR or pdf-parse) |
| `ministry` | `string` | Originating ministry (`interior`, `exteriores`, `defensa`) |
| `archive_id` | `string` | Archival signature (e.g., `AGA-83-07633`, `AGMAE-R39017`) |

## Dataset Statistics

| Metric | Value |
|---|---|
| **Total rows (pages)** | 1,224 |
| **Source documents** | 167 |
| **Text coverage** | 100% (all pages have extracted text) |
| **Total extracted text** | ~12.6 million characters |
| **File types** | 155 PDFs (92.8%), 12 JPGs (7.2%) |

### Documents by Ministry

| Ministry | Documents | Description |
|---|---|---|
| `defensa` | 108 | Military intelligence reports (CESID/CNI), internal memos |
| `exteriores` | 31 | Diplomatic cables, embassy communications, verbal notes |
| `interior` | 28 | Wiretap transcripts from the Guardia Civil, phone intercepts between conspirators |

### Document Characteristics

| Property | Breakdown |
|---|---|
| **Format** | 90 typed, 73 scanned, 4 handwritten |
| **Classification level** | 88 *secreto*, 77 *reservado*, 2 *confidencial* |
| **Pages per document** | Min: 1, Max: 312, Mean: 7.3 |
| **Text length per document** | Min: 130 chars, Median: 26.8K chars, Max: 3.45M chars |

## Data Processing Pipeline

### Image Rendering
- **Tool:** `pdf2image` + Poppler
- **Resolution:** 150 DPI (PNG)
- **JPG documents** are loaded directly as single-page images

### Text Extraction
Two extraction methods were used depending on document type:

1. **pdf-parse** (typed documents): Direct text extraction from digitally-created PDFs
2. **Vision-Language OCR** (scanned/handwritten documents):
   - **Model:** [`mlx-community/GLM-OCR-bf16`](https://huggingface.co/mlx-community/GLM-OCR-bf16) running on Apple Silicon via MLX
   - **Input resolution:** 200 DPI
   - **Max tokens per page:** 8,192
   - **Post-processing:** Whitespace normalization, paragraph structure preservation

The `isOcrText` field in the source metadata indicates which method was used per document.

## Usage

```python
from datasets import load_dataset

ds = load_dataset("JorgeAV/23f-expediente")

# Display a page
print(ds["train"][0]["text"])
ds["train"][0]["image"].show()

# Filter by ministry
interior = ds["train"].filter(lambda x: x["ministry"] == "interior")

# Get all pages from a specific document
doc_pages = ds["train"].filter(lambda x: x["pdf_source"] == "interior/guardia-civil/23F_1_conversacion_gc_tejero.pdf")
```

## Intended Uses

- **Historical research** on the Spanish Transition to democracy
- **Document AI** benchmarking: OCR, layout analysis, and document understanding on Spanish historical documents
- **NLP in Spanish**: text extraction quality evaluation, named entity recognition on official/archival language
- **Educational projects** about 20th-century European political history

## Limitations

- OCR quality is lower on handwritten documents (4 out of 167)
- Some scanned pages may contain text artifacts or recognition errors
- Text extraction from multi-column layouts may not preserve reading order perfectly
- All text is in Spanish; no translations are provided

## Source & License

- **Original documents:** Public domain (declassified records from the Spanish Government, published via the *Portal de Archivos Estatales*)
- **Structured dataset:** [CC-BY-4.0](https://creativecommons.org/licenses/by/4.0/)

## Associated Project

This dataset powers [23f-expediente](https://github.com/JorgeAV/23f-expediente), an interactive educational web application with an FBI/spy case-file aesthetic for exploring the 23-F documents.