--- license: apache-2.0 language: - km library_name: transformers pipeline_tag: translation tags: - transliteration - romanization - romanized-khmer - khmer - character-level - onnx - int8 - safetensors - keyboard - on-device base_model: samdei/samdei-type-mini-pretrained base_model_relation: finetune datasets: - samdei/samdei-type-pretraining-data - samdei/samdei-type-finetuning-data metrics: - cer - exact_match model-index: - name: Samdei Type Mini results: - task: type: translation name: Romanized Khmer to Khmer script dataset: type: samdei/samdei-type-bench name: Samdei Type Bench split: test metrics: - type: cer value: 0.058 name: CER (beam 5, word-LM rerank, 8-word pieces) - type: exact_match value: 0.239 name: Exact match --- # Samdei Type Mini Samdei Type Mini converts romanized Khmer, as people type it in chat, into Khmer script. It is a 7.1M-parameter character-level encoder-decoder, small enough for a phone keyboard: 7 MiB as int8 ONNX. | Input | Output | |---|---| | `knhom jong tov phsar` | ខ្ញុំចង់ទៅផ្សារ | | `ot mean wifi te` | អត់មាន wifi ទេ | Release article: [Samdei Type: open models that turn romanized Khmer into Khmer script](https://huggingface.co/blog/mengsay/samdei-type), with a demo of this model in a chat. ## Quick start With transformers alone (`pip install transformers torch`): ```python from transformers import AutoModelForSeq2SeqLM, AutoTokenizer repo = "samdei/samdei-type-mini" tok = AutoTokenizer.from_pretrained(repo) model = AutoModelForSeq2SeqLM.from_pretrained(repo) def pieces(text: str, max_words: int = 8) -> list[str]: """At most max_words words per piece, with a break after punctuation, as in the paper.""" out, cur = [], [] for w in text.split(): cur.append(w) if len(cur) == max_words or w[-1] in ",.;:?!។": out.append(" ".join(cur)) cur = [] if cur: out.append(" ".join(cur)) return out or [text] def convert(text: str) -> str: batch = tok(pieces(text), padding=True, truncation=True, return_tensors="pt") out = model.generate(**batch) # beam search, k = 5 (generation_config.json) return " ".join(tok.batch_decode(out, skip_special_tokens=True)) print(convert("knhom jong tov phsar")) # ខ្ញុំចង់ទៅផ្សារ ``` This is beam search alone (CER 0.072 on Samdei Type Bench). For the best quality, rerank with the word LM of this repo (0.058), using the [`samdei-type`](https://github.com/Samdei-Labs/samdei-type) package: ```bash pip install "samdei-type[deploy]" ``` ```python from samdei_type.pipeline import Converter conv = Converter("samdei/samdei-type-mini") # int8 ONNX + word-LM rerank, as in the paper print(conv.convert("knhom jong tov phsar")) # ខ្ញុំចង់ទៅផ្សារ ``` Or on the command line: `samdei-type "knhom jong tov phsar"`. ## Files ``` config.json, generation_config.json, model.safetensors (28.4 MiB), tokenizer*.json fp32 model (transformers) onnx/encoder_kv.int8.onnx (5.0 MiB), onnx/decoder_step.int8.onnx (2.0 MiB), onnx/vocab.json int8 ONNX lm/wordfull_nowiki_p25.json.gz (1.2 MiB) word bigram LM for reranking rerank.json decoding and rerank settings ``` **On a phone,** the two ONNX graphs are all a runtime needs. `encoder_kv` runs once per input. `decoder_step` decodes one token with a KV cache, and its batch axis holds all beams. [`samdei_type/onnx_infer.py`](https://github.com/Samdei-Labs/samdei-type) is a plain-numpy reference to port from. On one Apple M5 Pro CPU thread, beam 5 takes 4.9 ms per input (p50); with the rerank, a keystroke takes 5.7 ms. It has not been timed on a phone yet. ## Results [Samdei Type Bench](https://huggingface.co/datasets/samdei/samdei-type-bench): 2,155 everyday messages and words with human-checked Khmer. The metrics are CER after normalization and exact match. | System | CER | Exact | |---|---|---| | **Samdei Type Mini**, reranked | 0.058 | 24% | | Samdei Type Mini, beam only (the transformers example above) | 0.072 | 17% | | [Samdei Type Base](https://huggingface.co/samdei/samdei-type-base), reranked | 0.052 | 30% | | Samdei Type Prime (not released) | 0.027 | 61% | | Akkhara-mini, 6-word pieces (0.159 on whole inputs) | 0.066 | 23% | | AkaraAlpha | 0.281 | 2% | - Prime is Base trained further on scenario-designed, LLM-drafted, human-audited data. Most bench items share that scenario design, so its lead is an upper bound. - Decode long messages in pieces, as above. On whole inputs, beam-only CER is 0.218. ## Model and training - **Architecture.** Pre-LN Transformer encoder-decoder: 6 encoder and 2 decoder layers, d_model 256, 4 heads, FFN 1024. It has 215 input symbols (Latin letters, digits, punctuation) and 228 output symbols (Khmer characters, and copied characters). - **transformers format.** It loads as `BlenderbotForConditionalGeneration` because the architecture is the same; it has nothing else in common with BlenderBot. Both vocabularies share one id space (output symbols first), so transformers counts 7.43M parameters. - **Training.** Three stages on one laptop GPU (Apple M5 Pro), about 3 hours in all: 1. pretraining on 6.5M pairs: real Khmer sentences spelled by a rule-based typing generator ([`samdei-type-mini-pretrained`](https://huggingface.co/samdei/samdei-type-mini-pretrained)); 2. fine-tuning on 1.16M chat-like pairs; 3. adaptation to casual register (1.59M pairs, with replay). - **Data.** The exact training files are in [`samdei-type-pretraining-data`](https://huggingface.co/datasets/samdei/samdei-type-pretraining-data) and [`samdei-type-finetuning-data`](https://huggingface.co/datasets/samdei/samdei-type-finetuning-data), with `recipes.json`. No evaluation data and no other romanization system's data was trained on. Akkhara-Preview, a public 110-row preview, was used only to pick stage 1–2 checkpoints; stage 3 used a closed development set. ## Limitations - Every romanized input seen in training is synthetic (rules, mined spelling tables, an LLM). Real typing appears only in the evaluation. - Homophones that need more context than a word bigram: ដើរ / ដែរ (`der`), ដឹក / ដែក / ដេក (`dek`). - Loanwords typed in English (`check` for ឆែក) stay English. A Latin comma is copied, and a missing final ។ is never added. - Rare dictionary words without context are hard, as is heavier abbreviation than in training. - The Khmer side is mostly web text, with its topics and biases. - Not for Khmer → Latin, official romanization of names, or output nobody checks. ## License Apache-2.0 (`LICENSE`; keep `NOTICE` when redistributing). - **Word LM:** built without Wikipedia, from HPLT v2, FineWeb-2 (ODC-By), ALT (CC BY) and our LLM-written chat sentences. - **Training data:** CC BY-SA 4.0, with its sources in the dataset cards. - **Tools:** `khmercut` (Apache-2.0) and SIL `khnormal` (MIT). ## 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} } ```