README.md

---
dataset_info:
  features:
  - name: source
    dtype: string
  - name: tokens
    list: string
  - name: tags
    list: string
  splits:
  - name: train
    num_bytes: 273218408
    num_examples: 341968
  - name: test
    num_bytes: 14469157
    num_examples: 17999
  download_size: 268889691
  dataset_size: 287687565
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
  - split: test
    path: data/test-*
license: mit
task_categories:
- token-classification
language:
- la
size_categories:
- 100K<n<1M
---

# Latin ASR Post-Processing Dataset

A sequence-labeling dataset built for fine-tuning BERT-style models (e.g., `latin-bert`) on **Inverse Text Normalization (ITN)** — restoring capitalization and punctuation on raw, lowercased Latin text (such as [`njand/wav2vec2-xls-r-latin`](https://huggingface.co/njand/wav2vec2-xls-r-latin) ASR outputs).

Compiled from 2,141 files in the [CLTK Latin Library](https://github.com/cltk/lat_text_latin_library) and augmented with transcripts from the [`njand/llpsi-speech-dataset`](https://huggingface.co/datasets/njand/llpsi-speech-dataset) (currently private). Cleaned and transformed through a specialized classical Latin normalization pipeline.

---

## Quick Stats
* **Tokens:** ~11 million
* **Samples:** >350k lines
* **Task:** Token Classification / Sequence Labeling
* **Splits:** 95% Train (341,968 samples) / 5% Test (17,999 samples)

---

## Data Schema

Each example pairs lowercased, normalized tokens with a joint **Casing + Trailing Punctuation** target tag (14 possible combinations):

**Input Sentence:** “Dicam plane, Caesar, quod sentio.”

```json
{
  "source": "cicero/lig.txt",
  "tokens": ["dicam", "plane", "caesar", "quod", "sentio"],
  "tags": ["TITLE_NONE", "LOWER_COMMA", "TITLE_COMMA", "LOWER_NONE", "LOWER_PERIOD"]
}

```

* **Casing Labels:** `TITLE`, `LOWER`
* **Punctuation Labels:** `NONE`, `PERIOD` (`.`), `COMMA` (`,`), `COLON` (`:`), `SEMICOLON` (`;`), `EXCLAMATION` (`!`), `QUESTION` (`?`)

---

## Data Preprocessing Pipeline

1. **Filtering & Noise Removal:**
* Strips editorial headnotes, metadata, apparatus criticus, dates, section numbers, bracketed markers, and English notes.
* Drops incomplete sentence fragments, lines containing ALL-CAPS words, or unparseable lines.

2. **Praenomina Expansion:**
* Abbreviated Roman praenomina (e.g., `M.`, `C.`, `Sex.`, `Agr.`, etc.) are contextually expanded and inflected to agree grammatically with the following word (e.g., `M. Tullio` → `Marco Tullio`).

3. **Dynamic Roman Numeral Expansion:**
* Valid, grammatically indeclinable Roman numerals are dynamically converted to Latin cardinal words (e.g., `XIV` → `quattuordecim`), matching original casing.
* Vocabulary collisions with real Latin words (e.g., `i`, `vi`) are safely preserved.

4. **Classical Orthography (i/u Normalization):**
* Strips all macrons and diacritics.
* Expands ligatures (`æ`/`œ` → `ae`/`oe`).
* Normalizes orthography to strict classical standards: converts consonantal *v* to *u* and *j* to *i* (e.g., `uva` → `uua`, `jam` → `iam`).
* Normalizes compound *-iacere* verb forms (e.g., `ejicio` → `eicio`, `conjicere` → `conicere`).

5. **Sentence Splitting & Probabilistic Merging:**
* Sentence boundaries are tokenized using NLTK.
* Adjacent sentences within paragraphs are probabilistically merged (`p = 0.50`, max 1,000 characters) to simulate longer context windows.