samdei-type-base / README.md
mengsay's picture
Link the release article
06a3e56 verified
|
Raw History Blame Contribute Delete
7.73 kB
metadata
license: apache-2.0
language:
  - km
library_name: transformers
pipeline_tag: translation
tags:
  - transliteration
  - romanization
  - romanized-khmer
  - khmer
  - character-level
  - pytorch
  - onnx
  - int8
  - safetensors
base_model: samdei/samdei-type-base-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 Base
    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: 0.052
            name: >-
              CER (beam 5, char LM + word LM + OOV + channel rerank, 8-word
              pieces)
          - type: exact_match
            value: 0.303
            name: Exact match

Samdei Type Base

Samdei Type Base converts romanized Khmer, as people type it in chat, into Khmer script. It is an 18.2M-parameter character-level encoder-decoder, the most accurate open Samdei Type model, for desktop and server use. For a phone, use Samdei Type Mini.

Input Output
knhom jong tov phsar αžαŸ’αž‰αž»αŸ†αž…αž„αŸ‹αž‘αŸ…αž•αŸ’αžŸαžΆαžš
ot mean wifi te αž’αžαŸ‹αž˜αžΆαž“ wifi αž‘αŸ

Release article: Samdei Type: open models that turn romanized Khmer into Khmer script.

Quick start

With transformers alone (pip install transformers torch):

from transformers import AutoModelForSeq2SeqLM, AutoTokenizer

repo = "samdei/samdei-type-base"
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 ",.;:?!αŸ”":
            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"))  # αžαŸ’αž‰αž»αŸ†αž…αž„αŸ‹αž‘αŸ…αž•αŸ’αžŸαžΆαžš

This is beam search alone (CER 0.062 on Samdei Type Bench). For the best quality, rerank with the two LMs and the channel model of this repo (0.052), using the samdei-type package:

pip install samdei-type
from samdei_type.pipeline import Converter

conv = Converter("samdei/samdei-type-base")   # PyTorch weights + full rerank, as in the paper
print(conv.convert("knhom jong tov phsar"))   # αžαŸ’αž‰αž»αŸ†αž…αž„αŸ‹αž‘αŸ…αž•αŸ’αžŸαžΆαžš

Or on the command line: samdei-type -m samdei/samdei-type-base "knhom jong tov phsar".

Files

config.json, generation_config.json, model.safetensors (71.1 MiB), tokenizer*.json   fp32 model (transformers)
onnx/encoder_kv.int8.onnx (11.3 MiB), onnx/decoder_step.int8.onnx (6.3 MiB), onnx/vocab.json   int8 ONNX
channel/ (28.4 MiB)                        reverse model (Khmer -> romanized) for reranking, same format
lm/char5_nowiki.npz (8.9 MiB)              character 5-gram LM
lm/wordfull_nowiki.json.gz (12.2 MiB)      word bigram LM
rerank.json                                decoding and rerank settings

Reranking scores each beam hypothesis y for input x as log P(y|x) + 0.3 log P_char(y) + 0.3 log P_word(y) βˆ’ 1.0 #OOV(y) + 0.3 log P_ch(x|y). The channel term checks that y explains all of the input, so the LMs cannot swap a rare word for a frequent one. Nothing in this repo is a pickle.

Results

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 Base, reranked 0.052 30%
Samdei Type Base, beam only (the transformers example above) 0.062 24%
Samdei Type Mini, reranked 0.058 24%
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.
  • Without punctuation, Base's CER is 0.046. Decode long messages in pieces, as above. On whole inputs, beam-only CER is 0.133.
  • On one Apple M5 Pro CPU thread with int8 ONNX, beam 5 takes 9.3 ms per input (p50).

Model and training

  • Architecture. Pre-LN Transformer encoder-decoder: 6 encoder and 3 decoder layers, d_model 384, 6 heads, FFN 1536. It has 216 input and 228 output symbols. The channel model is Mini-sized (7.1M) and reads in the reverse direction.

  • 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 18.63M parameters.

  • Training. Three stages on one laptop GPU (Apple M5 Pro), about 6 hours in all:

    1. pretraining on 12.8M pairs: real Khmer sentences spelled by a rule-based typing generator (samdei-type-base-pretrained);
    2. fine-tuning on 1.16M chat-like pairs;
    3. adaptation to casual register (1.59M pairs, with replay).

    The rerank LMs are built from HPLT v2, FineWeb-2, ALT and our LLM-written chat sentences, without Wikipedia.

  • Data. The exact training files are in samdei-type-pretraining-data and 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 / αžŠαŸ‚αžš (der), αžŠαžΉαž€ / αžŠαŸ‚αž€ / αžŠαŸαž€ (dek).
  • Loanwords typed in English (check for αž†αŸ‚αž€) stay English. A Latin comma is copied, and a missing final αŸ” is never added.
  • Rare dictionary words without context are hard, many of them Pali-Sanskrit.
  • The Khmer side is mostly web text, with its topics and biases.
  • Not for Khmer β†’ Latin, official romanization of names, or output nobody checks.

License

Apache-2.0 (LICENSE; keep NOTICE when redistributing).

  • LMs: 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

@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}
}