Netra-Lab/Netra-NMT

Netra NMT

A 77M-parameter English↔Khmer translation model: one encoder-decoder transformer for both directions, trained from scratch on 220 million tokens of parallel text. It runs on ONNX Runtime with the netra-nmt package, on a CPU, without PyTorch.

Code: github.com/netra-ai-lab/Netra-NMT, with the inference package and the full training recipe.

Results

ALT test split (mutiyama/alt, 1,018 pairs), greedy decoding; spBLEU with the FLORES-200 tokenizer, higher is better.

Model	Params	EN → KM	KM → EN
NLLB-350M (EN-KM)	350M	18.89	0.26
NLLB-200 distilled 600M	600M	9.24	31.28
NLLB-200 distilled 1.3B	1.3B	16.62	35.15
MADLAD-400	3B	16.19	36.13
Google Translate	—	28.93	34.86
Netra NMT	77M	25.83	19.47

• Where it wins: English→Khmer, the best spBLEU after Google Translate and the best TER of all systems (60.47), and on CPU 65.6 ms per sentence greedy against 550.8 for NLLB-350M.
• Where it loses: Khmer→English, 15.7 spBLEU behind NLLB-200 1.3B.

spBLEU, chrF++, TER, COMET (wmt22-comet-da) and BERTScore F1 (xlm-roberta-base) on ALT, both directions.

English→Khmer, median ms per sentence over 20 sentences, Intel i7-13700KF (16 threads) and RTX 3060, PyTorch.

Architecture

• Shape: 12 encoder layers, 2 decoder layers (deep encoder, shallow decoder), d_model 512, 8 heads, FFN 2048, vocabulary 32K.
• Components: rotary position embeddings, RMSNorm (pre-norm), SwiGLU, one embedding table shared by the encoder, the decoder and the tied output head, a KV cache.
• Input: a target-language marker (<2km> or <2en>) then the source. English sources are NFC-normalised and lowercased; Khmer is word-segmented with khmercut. The package does this for you.

How it was made

1. Data. About 2.4M unique pairs from four sources (synthetic and ParaCrawl), deduplicated, length-filtered, then used in both directions: 4.2M examples.
2. Tokenizer. A 32K SentencePiece unigram model shared by both languages, trained on Khmer segmented with khmercut.
3. Training. From scratch, with label smoothing and a warm-up + cosine learning-rate schedule (settings in training/netra_nmt_train/scripts/recipe/09_train_student.py); the checkpoint with the best validation chrF is released.
4. Export. The encoder and a one-step decoder with a KV cache as ONNX graphs, weights stored as fp16. The exported folder generates the same tokens as the PyTorch model on the ALT test set (greedy and beam-5, both directions).

Usage

pip install "netra-nmt[cpu]"        # or [gpu]

from netra_nmt import translate

translate("Hello, how are you?")                         # English -> Khmer
translate("ខ្ញុំស្រលាញ់ប្រទេសកម្ពុជា។", direction="km2en")    # -> "I love Cambodia."

netra-nmt --text "Hello, how are you?" --mode beam
netra-nmt serve                       # web app + REST API (extra: server)

The package downloads model/ from this repo once, at the commit it was tested with.

Files

File	Contents
model/netra-model.json	settings: graphs, decoder shape, special ids, language markers, sha256 of every file
model/encoder.onnx	tokens → cross-attention keys/values for every decoder layer (136 MB)
model/decoder_step.onnx	one decoding step with the self-attention KV cache (52 MB)
model/spm_32k.model	the SentencePiece tokenizer
pytorch/model.safetensors	fp16 weights for fine-tuning (lm_head tied to the embeddings, not stored)
pytorch/config.json, pytorch/spm_32k.model	architecture and tokenizer for the PyTorch model
assets/	the figures on this card

To fine-tune from pytorch/: bash training/finetune.sh my_pairs.tsv in the GitHub repo.

License and attribution

• Weights: MIT, trained from scratch (no pretrained weights).
• Training data: see the dataset card and its sources; check their terms before commercial use.
• Evaluation data: ALT (NICT), CC BY 4.0.