---
language: [km]
license: cc-by-4.0
task_categories: [text-generation]
tags: [khmer, spell-checking]
---
# Khmer spell-checking pairs (clean v2)

Cleaned version of `phonsobon/khmer-spellcheck-pairs-clean`, built by the Khmer spell-checker training notebook.

Columns: `source` (text with errors), `target` (correct text), `origin` (where the clean text came from).

## Splits
| Split | Rows |
|---|---|
| train | 1,588,249 |
| test_synthetic | 8,102 |
| validation | 7,385 |
| test_real | 51,915 |

- `test_real` contains **only real human errors** (seanghay/khmer-spelling-corrections and
  S-Sethisak/khmer-orthography-correction-dataset). They are **not** in train or validation.
- No target sentence appears in two splits.

## Cleaning
Khmer only (any other script removed), repeated-character bug, length and edit limits
(synthetic rows), broken Khmer character order, known misspellings in targets, duplicates,
leakage between splits. Targets use the RAC form ឱ្យ.

| split          |   original |    kept |   invalid_khmer |   too_many_edits |   repeat_bug |   length_ratio |
|:---------------|-----------:|--------:|----------------:|-----------------:|-------------:|---------------:|
| train          |    2047228 | 2037930 |            7911 |             1147 |          167 |             73 |
| test_synthetic |       9643 |    9624 |               0 |               12 |            0 |              7 |
| validation     |      10418 |   10406 |               9 |                0 |            3 |              0 |
| test_real      |       2632 |    2624 |               6 |                0 |            2 |              0 |