---
language:
  - en
license: mit
task_categories:
  - text-generation
size_categories:
  - n<1K
tags:
  - transliteration
  - diacritization
  - arabic
  - asr-correction
  - sermon
  - islamic-education
pretty_name: logical-transcripts
---

# logical-transcripts

Golden paired dataset for training models to transliterate Arabic Latin text into
scholarly diacritized form — built from a single recorded Islamic lecture
(Chapter 24, Lecture 16) with a raw ASR transcript and a human-polished scholarly
transcript.

Two artifacts are stored separately for provenance and review:

| File | Rows | Purpose |
|------|------|---------|
| `train.jsonl` | 203 | **Golden** — quality-filtered pairs for training |
| `bronze.jsonl` | 773 | **Bronze** — every aligned sentence pair before filtering, tagged with `category` + `diacritics` count for provenance/review |

## Golden quality criteria

Rows in `train.jsonl` meet **both**:

1. `input != output` (no identity rows)
2. Output contains **≥ 2 distinct diacritized letters** — each diacritic-carrying base
   letter counts once (`ā`, `ḥ`, `ṣ`, …), plus the `ʿ` / `ʾ` hamza-ʿayn spacing
   modifier letters.

Markdown asterisks from the source transcript are stripped from outputs.

## Schema

```json
{
  "instruction": "Transliterate the following Arabic Latin text to scholarly diacritized form:",
  "input": "There was no athan.",
  "output": "There was no aẓān."
}
```

The `instruction`/`input`/`output` schema follows the standard instruction-tuning
convention, so it can be concatenated with other transliteration datasets for
training.

## Task

The `input` is raw, un-diacritized Arabic-as-spoken-in-Latin-script (including ASR
artifacts: stutters, mis-heard words, run-on sentences). The `output` is the
scholarly diacritized transliteration (macrons, sub-dots, hamza/ʿayn, word
corrections, cleaned punctuation). Rows therefore train **diacritization + ASR
correction** jointly — a broader task than a clean-input transliteration baseline.

## Statistics

- 203 golden rows, 773 bronze rows
- Golden input length: median 90 chars, max 1737
- Diacritic-letter distribution: 2×85, 3×53, 4×22, 5×23, 6×8, 7×9, 8×3

## Provenance

Source material: 8 `{input_text, output_text}` chunk pairs extracted from a single
recorded Islamic lecture (Chapter 24, Lecture 16). The concatenated inputs exactly
reconstruct the raw ASR transcript; the concatenated outputs exactly reconstruct the
polished scholarly transcript. Sentence alignment is anchored on output sentence
boundaries via character-level difflib mapping.

## Reproduction

```bash
python3 scripts/convert_golden.py
```

Writes `train.jsonl` (golden) and `bronze.jsonl`.