---
language: [si, pi, sa]
license: other
task_categories: [text-classification]
pretty_name: Sinhala-script LangID (Sinhala, Pali, Sanskrit)
configs:
  - config_name: default
    data_files:
      - {split: train, path: train.jsonl}
      - {split: validation, path: validation.jsonl}
      - {split: test, path: test.jsonl}
---

# Sinhala-script language identification: Sinhala, Pali, Sanskrit

Sentence-level instances of three languages written in Sinhala script, with a
leakage-free document-blocked split. Labels: `sin_Sinh`, `pli_Sinh`, `san_Sinh`.

## Rows

| split | sin_Sinh | pli_Sinh | san_Sinh | total |
|---|---:|---:|---:|---:|
| train | 83105 | 58309 | 10565 | 151979 |
| validation | 10345 | 7317 | 1337 | 18999 |
| test | 10567 | 7105 | 1310 | 18982 |

## Construction

Built by `scripts/maintainer/resplit_target.py` (config: `target_split` in
`config/pipeline.yaml`):

1. Sources pooled and NFC-normalised; text not in Sinhala script removed.
2. Documents: source document ids; for the Pali-Sinhala parallel corpus
   (sinhala-nlp/pali-sinhala, row-aligned) a sutta starts at each Pali incipit
   *evam me sutam*, and both sides of a translation pair share the document.
3. Split unit: contiguous block of <= 25 rows within a document.
4. Sentence-level units: split at sentence terminators, packed to <= 200 characters.
5. Exact duplicates removed; texts with conflicting labels removed.
6. Near duplicates removed with MinHash LSH (character 5-grams, Jaccard >= 0.8).
7. Group-stratified {'train': 0.8, 'validation': 0.1, 'test': 0.1} split (seed 42).

Leakage check: {'exact_cross_split': 0, 'near_dup_cross_split_j0.8': 0, 'near_dup_cross_split_j0.5': 15005}

Length (characters): {'pli_Sinh': {'p50': 138.0, 'p90': 195.0, 'max': 200}, 'san_Sinh': {'p50': 108.0, 'p90': 173.0, 'max': 200}, 'sin_Sinh': {'p50': 155.0, 'p90': 197.0, 'max': 200}}

## Sources

- DCS: {'san_Sinh': 4888}
- SansinNT: {'san_Sinh': 8223}
- SiDiaC-v2: {'san_Sinh': 101, 'pli_Sinh': 2862}
- SiPaKosa: {'sin_Sinh': 5997}
- pali-sinhala-parallel: {'pli_Sinh': 69869, 'sin_Sinh': 98020}

## Fields

`sample_id`, `text`, `label`, `source`, `subcorpus`, `doc_id`, `block_id`, `text_sha256`.