---
dataset_info:
  features:
  - name: plaintext
    dtype: string
  - name: plaintext_with_errors
    dtype: string
  - name: ciphertext_with_errors_with_mix_code
    list: string
  - name: ciphertext_with_errors_without_mix_code
    list: string
  - name: ciphertext_without_errors_with_mix_code
    list: string
  - name: ciphertext_without_errors_without_mix_code
    list: string
  - name: key_with_errors_with_mix_code
    dtype: string
  - name: key_with_errors_without_mix_code
    dtype: string
  - name: key_without_errors_with_mix_code
    dtype: string
  - name: key_without_errors_without_mix_code
    dtype: string
  - name: year
    dtype: int64
  - name: year_range
    dtype: string
  - name: text_length
    dtype: int64
  - name: text_id
    dtype: string
  - name: plaintext_with_errors_with_mix
    dtype: string
  - name: plaintext_with_errors_without_mix
    dtype: string
  splits:
  - name: train
    num_bytes: 3001704620
    num_examples: 351566
  - name: validation
    num_bytes: 374866413
    num_examples: 43926
  - name: test
    num_bytes: 375564270
    num_examples: 43966
  download_size: 1445078159
  dataset_size: 3752135303
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
  - split: validation
    path: data/validation-*
  - split: test
    path: data/test-*
license: cc-by-4.0
language:
- es
tags:
- historical
- cryptanalysis
- cipher
- decipherment
- historical-cryptology
- substitution-cipher
- homophonic-cipher
- homophonic-substitution
- low-resource
- siachronic
- synthetic-data
size_categories:
- 100K<n<1M
---

# Dataset Card for HistCiph — Spanish

## Table of Contents
- [Dataset Description](#dataset-description)
  - [Dataset Summary](#dataset-summary)
  - [Supported Tasks and Leaderboards](#supported-tasks-and-leaderboards)
  - [Languages](#languages)
- [Dataset Structure](#dataset-structure)
  - [Data Instances](#data-instances)
  - [Data Fields](#data-fields)
  - [Data Splits](#data-splits)
- [Dataset Creation](#dataset-creation)
  - [Curation Rationale](#curation-rationale)
  - [Source Data](#source-data)
  - [Annotations](#annotations)
- [Considerations for Using the Data](#considerations-for-using-the-data)
  - [Social Impact of Dataset](#social-impact-of-dataset)
  - [Discussion of Biases](#discussion-of-biases)
  - [Other Known Limitations](#other-known-limitations)
- [Additional Information](#additional-information)
  - [Dataset Curators](#dataset-curators)
  - [Licensing Information](#licensing-information)
  - [Citation Information](#citation-information)

## Dataset Description

### Dataset Summary

The Spanish subset of **HistCiph** is part of the first publicly available multilingual collection of historically grounded plaintext–ciphertext pairs for classical homophonic substitution ciphers. It pairs diachronically balanced historical Spanish plaintext with independently generated homophonic substitution keys and controlled transcription noise, producing four distinct ciphertext variants per plaintext document.

The dataset is designed to support reproducible research on neural cryptanalysis, character-level sequence modelling, cross-lingual transfer, and decipherment. Synthetic generation is explicitly constrained by documented properties of historical ciphers — including multi-homophone allocation, variable-length numeric codes, and transcriptional irregularities — and validated using information-theoretic diagnostics (entropy, redundancy, frequency masking, and unicity distance).

### Supported Tasks and Leaderboards

- `text2text-generation`: The dataset can be used to train and evaluate models for **automatic decipherment** — recovering plaintext from ciphertext. Models map sequences of numeric ciphertext tokens (3- or 4-digit codes) back to character-level plaintext.
- `translation`: The aligned plaintext–ciphertext pairs can be framed as a translation task between a natural-language plaintext "source" and a ciphertext "target" (or vice versa), supporting research on alignment and key extraction.
- `cipher-key-recovery`: Predicting the homophonic substitution key from ciphertext (or plaintext–ciphertext pairs).
- `cipher-type-classification`: Distinguishing between fixed-length and variable-length code regimes, and between clean and noisy transcription regimes.
- `robustness-evaluation`: Measuring model robustness to stochastic noise simulating archival transcription errors.

No active leaderboard is maintained for this dataset at this time.

### Languages

The dataset is monolingual Spanish (BCP-47: `es`), written in the Latin script. It captures both pre-standardization and post-standardization stages of the language, with diachronic coverage spanning multiple centuries (approximately 1500–1799).

Historical Spanish text in the corpus reflects period-specific orthographic conventions, including spellings, ligatures, and graphemic variants that differ substantially from modern norms. This orthographic variation is preserved by design and is integral to the dataset's research value.

## Dataset Structure

### Data Instances

Each instance is a structured record containing three plaintext variants, four corresponding ciphertext variants (mixed/fixed code length × with/without transcription errors), four matching homophonic substitution keys, two plaintext alignment views, and metadata. A truncated example:

```json
{
  "plaintext": "[clean normalized plaintext, e.g. 'THEGOSPELOFOURLORD...']",
  "plaintext_with_errors_with_mix": "[plaintext with '#' markers at noise positions, aligned with mixed-code+errors ciphertext]",
  "plaintext_with_errors_without_mix": "[plaintext with '#' markers at noise positions, aligned with fixed-code+errors ciphertext]",
  "ciphertext_with_errors_with_mix_code": "8148 0511 961 2209 ... 079",
  "ciphertext_with_errors_without_mix_code": "7533 0280 6091 3906 ... 4567",
  "ciphertext_without_errors_with_mix_code": "5903 1187 840 2910 ... 322",
  "ciphertext_without_errors_without_mix_code": "7592 3475 6912 6940 ... 0382",
  "key_with_errors_with_mix_code": {"A": ["284", "123"], "B": ["7876"], ...},
  "key_with_errors_without_mix_code": {"A": ["8824", "3381"], "B": ["1579"], ...},
  "key_without_errors_with_mix_code": {"A": ["081", "271"], "B": ["4428"], ...},
  "key_without_errors_without_mix_code": {"A": ["1499"], "B": ["6659"], ...},
  "year": 1568,
  "year_range": "1500-1599",
  "text_length": 50,
  "text_id": "text_50_..."
}
```

The four ciphertext variants are organized along two binary axes (code length × transcription noise), enabling controlled experimentation:

| | **Mixed (3+4 digit)** | **Fixed (4 digit)** |
|---|---|---|
| **With errors** | `ciphertext_with_errors_with_mix_code` | `ciphertext_with_errors_without_mix_code` |
| **Without errors** | `ciphertext_without_errors_with_mix_code` | `ciphertext_without_errors_without_mix_code` |

Each ciphertext variant is paired with its corresponding key (same naming convention with `key_` prefix). For variants involving transcription noise, a matching plaintext variant marks affected character positions with `#`, preserving alignment.

### Data Fields

| Field | Type | Description |
|---|---|---|
| `plaintext` | string | Clean, normalized historical plaintext (whitespace, formatting, and punctuation removed). Aligned with the two `ciphertext_without_errors_*` variants. |
| `plaintext_with_errors_with_mix` | string | Plaintext with `#` symbols inserted at positions affected by transcription noise. Aligned with `ciphertext_with_errors_with_mix_code`. |
| `plaintext_with_errors_without_mix` | string | Plaintext with `#` markers, aligned with `ciphertext_with_errors_without_mix_code`. |
| `ciphertext_with_errors_with_mix_code` | list | Ciphertext with stochastic insertion/deletion noise (p = 0.05) and variable-length 3/4-digit codes. |
| `ciphertext_with_errors_without_mix_code` | list | Ciphertext with transcription noise and fixed-length 4-digit codes. |
| `ciphertext_without_errors_with_mix_code` | list | Clean ciphertext with variable-length 3/4-digit codes. |
| `ciphertext_without_errors_without_mix_code` | list | Clean ciphertext with fixed-length 4-digit codes. |
| `key_with_errors_with_mix_code` | dict | Homophonic substitution key (character → list of 1–5 numeric codes) used for `ciphertext_with_errors_with_mix_code`. |
| `key_with_errors_without_mix_code` | dict | Key for the fixed-length + errors variant. |
| `key_without_errors_with_mix_code` | dict | Key for the variable-length clean variant. |
| `key_without_errors_without_mix_code` | dict | Key for the fixed-length clean variant. |
| `year` | int | Year of composition of the plaintext source. |
| `year_range` | string | Century-level bin (e.g. `"1500-1599"`) for temporal filtering and stratification. |
| `text_length` | int | Plaintext length bin in characters (one of: 50, 100, 200, 400, 600, 800, 1000). Actual length may be up to -10/+50 characters shorter/longer than the bin value. |
| `text_id` | string | Unique document identifier. |

Keys are generated independently per document — no key is reused across instances. For ciphertext variants with transcription noise, the realized ciphertext length may exceed the corresponding plaintext length due to insertions (up to ~50 additional codes for the longest texts).

### Data Splits

Each plaintext document carries four ciphertext variants and four corresponding keys; ciphertext counts above reflect the total number of ciphertext sequences across all variants.

Documents are stratified across plaintext length bins (50, 100, 200, 400, 600, 800, 1000 characters) and across 100-year intervals to support diachronic and length-based experimentation.

### Dataset Statistics

The following empirical statistics characterize the Spanish subset and were used to validate that the synthetic encryption procedure preserves the cryptographic properties of historical homophonic ciphers. Reported in Bruton et al. (2026).

**Plaintext characteristics**

| Statistic | Value |
|---|---:|
| Total plaintext characters | 74,288,880 |
| Unique character inventory | 138 |
| Shannon entropy (bits/char) | 4.177 |
| Theoretical max entropy (bits/char) | 7.109 |
| Redundancy (bits/char) | 2.932 |
| Rare-character count (cumulative mass < 0.005%) | 85 |
| Rare-character cumulative mass | 7.98e-05 |

**Ciphertext characteristics** (averaged across the four ciphertext variants)

| Statistic | Value |
|---|---:|
| Ciphertext entropy (bits/token) | 13.361 |
| Theoretical max entropy (bits/token) | 13.425 |
| % of theoretical maximum | 99.52% |

The ciphertext entropy approaching the theoretical upper bound indicates that homophonic substitution effectively masks plaintext frequency imbalances, producing near-uniform ciphertext token distributions independent of the underlying language.

**Homophonic key characteristics** (averaged across the four key variants)

| Statistic | Value |
|---|---:|
| Mean homophones per character | 2.87 |
| Median homophones per character | 3.0 |
| Min homophones per character | 1 |
| Max homophones per character | 5 |

Homophone allocation is frequency-dependent within each text: frequent characters receive up to 5 codes, rare characters receive 1.

**Cryptographic recoverability**

| Statistic | Value |
|---|---:|
| Mean unicity distance (chars) | 8.73 |

Unicity distance `U = H(K) / R` is the expected ciphertext length required to uniquely determine the encryption key. Because keys are generated independently per document and homophone allocation scales with text length, realized unicity distance grows monotonically with plaintext length. Most ciphertexts in this dataset exceed their corresponding unicity distance, placing them in the theoretically uniquely solvable regime.

## Dataset Creation

### Curation Rationale

Historical ciphertext decryption is an extreme low-resource setting: aligned plaintext–ciphertext pairs are scarce, often institutionally restricted, and unevenly distributed across languages and periods. Existing publicly available cipher datasets predominantly use modern plaintext, lack diachronic metadata, and/or embed ciphertexts within prompt templates rather than providing aligned sequence pairs. Prior work on neural cryptanalysis routinely generates synthetic data internally without releasing it, severely limiting reproducibility.

HistCiph addresses this gap by combining (i) diachronically balanced historical plaintext, (ii) independently generated homophonic substitution keys, and (iii) controlled transcription noise — all under explicit constraints derived from documented historical encryption practice. The resulting datasets enable systematic experimentation under conditions that approximate real archival material while supporting reproducible benchmarking.

### Source Data

#### Initial Data Collection and Normalization

Plaintext data was sourced from the Spanish subcorpora of [HistCorp](https://cl.lingfil.uu.se/histcorp/), a large-scale collection of historical texts covering multiple languages, time periods, and genres (Pettersson and Megyesi, 2018).

Prior to encryption, all plaintext data underwent the following normalization:
1. Removal of whitespace, formatting characters, and punctuation.
2. Filtering to minimize the presence of extended foreign-language passages, maintaining as close to monolingual content as possible.
3. Approximately equal sampling across 100-year intervals to support diachronic balance.
4. Length binning into seven categories (50, 100, 200, 400, 600, 800, 1000 characters).

Ciphertexts were then generated from the normalized plaintext using the [ChronoFidelius](https://github.com/mbruton0426/ChronoFidelius) toolkit, which implements the homophonic substitution and noise injection procedures described below.

**Ciphertext generation procedure:** For each plaintext character `c ∈ Σ`, a homophone set `H_c ⊂ Γ³ ∪ Γ⁴` is sampled, where `1 ≤ |H_c| ≤ 5` and each element is a unique 3- or 4-digit numeric code drawn from a global pool of 11,000 possible sequences. Encryption is defined as `E(c) ~ Uniform(H_c)`: each occurrence of `c` is replaced by a code sampled uniformly at random from its homophone set. Keys are generated independently for each document.

In variants including transcription noise, after each homophonic code is sampled an error operation is applied with probability `p = 0.05`. Errors are either deletions (the sampled ciphertext token is removed) or insertions (an additional token is sampled uniformly from tokens already present in the current ciphertext). At most one error is applied per plaintext position. Affected plaintext positions are marked with `#` in the corresponding plaintext variant to preserve alignment information.

#### Who are the source language producers?

The plaintext was produced by historical authors of Spanish texts spanning approximately 1500–1799. The originating documents reflect a range of historical genres and registers as available in HistCorp; specific producer demographics are not consistently recoverable from the source materials. The ciphertext, keys, and noise-marked plaintext variants are machine-generated by the ChronoFidelius toolkit.

### Annotations

The dataset does not contain manual annotations beyond the original source text from HistCorp. All ciphertext variants, keys, and alignment markers are deterministically or stochastically generated by the encryption pipeline. Plaintext metadata (year, year_range, text_length, text_id) is automatically derived from HistCorp source records and from the generation pipeline.

## Considerations for Using the Data

### Social Impact of Dataset

This dataset is intended to support digital humanities, historical, and cryptanalytic research by enabling the development of automatic decipherment and cryptanalysis methods that can assist scholars working with archival ciphered documents. Successful methods developed on HistCiph could substantially reduce the manual labour required to recover historical encrypted correspondence, opening up otherwise inaccessible primary sources for historians, philologists, and political historians.

The risks associated with this dataset are limited. Classical homophonic ciphers are cryptographically obsolete and have no contemporary security applications.

### Discussion of Biases

The plaintext reflects what was preserved, transcribed, and digitized from the historical Spanish record. As is typical of historical text collections, the available material is likely biased toward written registers produced by literate elites, institutional sources (church, state, learned societies), and male authors, with limited representation of vernacular, oral, or marginalized voices.

The encryption procedure itself is uniform across all texts and does not introduce additional sociolinguistic bias. However, models trained on this data may learn distributional patterns that reflect the orthographic and lexical conventions of the represented historical sources, and may generalize less well to underrepresented periods, registers, or regional varieties.

### Other Known Limitations

The dataset uses synthetic encryption of real historical plaintext rather than real plaintext–ciphertext pairs from archival sources. While this enables controlled experimentation and reproducible benchmarking, it does not capture all properties of authentic archival ciphers:

- **No semantic obfuscation.** Real historical ciphers often replaced sensitive names, places, or political terms with dedicated codebook entries, used null symbols to mislead frequency analysis, and compressed semantically salient expressions into single high-value codes. These practices are not modelled.
- **Independent keys per document.** Archival ciphers commonly reused keys across documents, sometimes over extended periods, creating cross-document statistical dependencies that can facilitate cryptanalysis. HistCiph generates independent keys per text to ensure clean experimental control.
- **Uniform homophonic allocation.** Authentic historical keys exhibit uneven allocation, special-purpose symbols, hierarchical groupings, and ad hoc expansions. The HistCiph encryption procedure uses controlled allocation to isolate the effects of redundancy, entropy, and code multiplicity.
- **Fixed transcription noise rate.** Errors are injected at a fixed rate of `p = 0.05` per character. Real archival error rates vary substantially with manuscript condition, scribal practice, and editorial conventions.
- **Long-tail character inventories.** Large character inventories may partly reflect residual loanwords and rare graphemic variants surviving normalization. Cumulative probability mass of these characters is negligible (< 0.005%) but can slightly inflate theoretical redundancy estimates.

## Additional Information

### Dataset Curators

The dataset was curated by Micaella Bruton (Stockholm University), Meriem Beloucif (Uppsala University), and Beáta Megyesi (Stockholm University).

Funding was provided by Riksbankens Jubileumsfond under grant **M24-0028**. Computational resources were provided by the National Academic Infrastructure for Supercomputing in Sweden (NAISS), partially funded by the Swedish Research Council under grant agreement **no. 2022-06725**.

Plaintext source data was originally compiled within HistCorp by Eva Pettersson and Beáta Megyesi (2018).

### Licensing Information

This dataset is released under the **Creative Commons Attribution 4.0 International License (CC BY 4.0)**.
See <https://creativecommons.org/licenses/by/4.0/> for the full license text.

### Citation Information

```bibtex
@inproceedings{bruton2026histciph,
  author    = {Bruton, Micaella and Beloucif, Meriem and Megyesi, Be\'ata},
  title     = {..},
  booktitle = {..},
  year      = {2026},
  note      = {Accepted LREC-COLING 2026/RESOURCEFUL. To appear. Citation information to be updated}
}
```