mengsay's picture
Link the technical report (PDF on GitHub) in the card and citation
11cdfcb verified
|
Raw History Blame Contribute Delete
6.86 kB
---
pretty_name: Samdei Type Resources
license: other # per config, see "License"
license_name: mixed-cc-by-4.0-cc0-apache-2.0
language:
- km
task_categories:
- translation
- text-generation
tags:
- transliteration
- romanization
- romanized-khmer
- khmer
- lexicon
- synthetic
size_categories:
- 1K<n<10K
multilinguality: monolingual
source_datasets:
- original
configs:
- config_name: chat_sentences
default: true
data_files:
- split: train
path: chat_sentences/chat_sentences.jsonl
- config_name: chat_sentences_romanized
data_files:
- split: train
path: chat_sentences/chat_sentences_romanized.jsonl
- config_name: chat_lexicon
data_files:
- split: train
path: lexicons/chat_lexicon.jsonl
- config_name: places
data_files:
- split: train
path: lexicons/places.jsonl
- config_name: english_chat_words
data_files:
- split: train
path: lexicons/english_chat_words.jsonl
- config_name: spelling_tables
data_files:
- split: train
path: spelling_tables/spelling_tables.jsonl
---
# Samdei Type Resources
These are the small reusable pieces behind the Samdei Type training data:
- **LLM-written Khmer chat sentences** (2,483), with Gemma 4 romanizations of most of them;
- **two spelling tables** mined from public lists of romanized Khmer words;
- **three hand-written lexicons:** chat spellings of frequent words, place names and English chat words.
They are inputs of our data generator, which is not public. The training pairs made from them are in [`samdei/samdei-type-pretraining-data`](https://huggingface.co/datasets/samdei/samdei-type-pretraining-data) and [`samdei/samdei-type-finetuning-data`](https://huggingface.co/datasets/samdei/samdei-type-finetuning-data).
Paper: [technical report](https://github.com/Samdei-Labs/samdei-type/blob/main/paper/samdei-type.pdf). Inference code: [github.com/Samdei-Labs/samdei-type](https://github.com/Samdei-Labs/samdei-type).
## Configs
| Config | Rows | Content | License |
|---|---|---|---|
| `chat_sentences` | 2,483 | Khmer chat sentences written by Gemma 4 12B: `khmer`, `domain` (50), `register` (4) | CC BY 4.0 |
| `chat_sentences_romanized` | 2,133 | Gemma 4 romanizations of those sentences: `khmer`, `latin`, `style` (0–3), `ok` (passed our filters) | CC BY 4.0 |
| `chat_lexicon` | 293 | chat spellings of 99 frequent words: `khmer`, `latin`, `weight`, `is_abbreviation` | CC0 |
| `places` | 76 | Cambodian provinces, Phnom Penh areas and landmarks: `khmer`, `latin_variants` (a list) | CC0 |
| `english_chat_words` | 358 | English and Latin tokens common in Khmer chat, copied verbatim on both sides: `word` | CC0 |
| `spelling_tables` | 368 | the two mined tables, flattened: `table`, `slot`, `key`, `spelling`, `prob` | Apache-2.0, with attribution |
The original JSON tables, which the generator reads, are in `spelling_tables/rules_chh.json` and `spelling_tables/rules_wd.json`.
## The pieces
**Chat sentences.** Web text is mostly news and prose. To cover chat topics, Gemma 4 12B, run locally, wrote Khmer chat sentences across 50 domains and 4 registers:
- casual chat between close friends;
- chat between a customer and a seller;
- a polite message from a younger person to an older one;
- a family group chat.
They are the only Khmer targets of the Samdei Type training data that are not human-written, and they were never used for evaluation. The file has 2,483 rows and 2,350 distinct texts.
**Romanizations.** Gemma 4 romanized the sentences "the way Cambodians type on Messenger", with 4 style prompts and 10 sentences per call. Each row's `khmer` is the sentence as it entered the training corpus: normalized, deduplicated, and split where a generated line held more than one sentence. It can therefore differ from the `chat_sentences` row it came from. `ok` marks the rows that passed our filters and became training pairs.
**Spelling tables.** These are per-slot tables (onset, vowel, coda, glide, context) of how people spell each Khmer sound in Latin letters. They were learned with Viterbi alignment and hard EM from two public lists:
- **`rules_chh`:** the Chhunneng word list, 24.9k (Khmer word, romanization) pairs in a consistent, dictionary-like scheme (ខ្ញុំ → `khnhom`, ឆ → `chh`). Held-out log-likelihood per letter: −1.31 with our hand tables, −0.86 with the mined table.
- **`rules_wd`:** 1.6k Wikidata person and place names, English label against Khmer label. This is the official, name-style scheme (Kampong, Banteay, Srae).
Words that are targets of a word-level evaluation set were removed from both lists before mining. Each `(table, slot, key)` row group gives a distribution over spellings, and `prob` sums to 1 within a group.
**Chat lexicon.** For 99 frequent words, it gives the spellings chat uses with their sampling weights, e.g. `knhom`, `khnhom`, `kyom` and `nh` for ខ្ញុំ, or `tv` for ទៅ. `is_abbreviation` marks the shortcuts. An abbreviation listed here is reserved for its word: the generator never produces it for another word.
## How to load
```python
from datasets import load_dataset
sents = load_dataset("samdei/samdei-type-resources", "chat_sentences", split="train")
lex = load_dataset("samdei/samdei-type-resources", "chat_lexicon", split="train")
```
## Limitations
- **The chat sentences are LLM output.** They can contain unnatural phrasing or spelling mistakes, and their topics follow the 50 domains we listed.
- **The Gemma romanizations** reflect one LLM's idea of chat spelling. They are looser and less regular than a rule-based generator, but they are not real typing.
- **Each spelling table reproduces its source's scheme:** a dictionary scheme for `rules_chh` and an official name scheme for `rules_wd`, not chat spelling as such.
- **The lexicons are small and hand-written** by the authors.
## License
| Piece | License | Attribution |
|---|---|---|
| Chat sentences and their romanizations | CC BY 4.0 | written with Gemma 4 (Apache-2.0) |
| Chat lexicon, places, English chat words | CC0 | – |
| `rules_chh` table | Apache-2.0 | derived from the Chhunneng romanized Khmer dataset, MIT License, © 2025 Chrea Chanchhunneng (notice below) |
| `rules_wd` table | Apache-2.0 | derived from Wikidata (CC0) |
The MIT notice of the Chhunneng dataset must accompany `rules_chh`. Its full text is included as `spelling_tables/LICENSE-chhunneng.txt`.
## Citation
```bibtex
@misc{samdei-type-2026,
title = {Samdei Type: Separating What Is Written from How It Is Spelled for Romanized Khmer Transliteration},
author = {Mengsay Loem},
year = {2026},
note = {Technical report},
url = {https://github.com/Samdei-Labs/samdei-type/blob/main/paper/samdei-type.pdf}
}
```
Contact: open a discussion in this repository's [Community tab](https://huggingface.co/datasets/samdei/samdei-type-resources/discussions).