kyleparrratt/luganda-translator-v1

Luganda Translator V1

High-performance English → Luganda translation. Outperforms Google Translate. 918 MB. Fully offline.

Benchmark (FLORES-200 devtest)	chrF++	spBLEU
Google Translate	41.29	14.45
Luganda Translator V1	45.01	17.70

Results: Statistically significant lead over Google Translate (paired bootstrap, p ≤ 0.001). Zero dependencies on APIs, internet, or GPUs. Optimized for CPU and Android (arm64).

Technical Specifications

• Performance: Superior to Google Translate on FLORES-200 benchmarks.
• Efficiency: 909M parameters (int8). 1.6× faster than the 1.37B source model.
• Dataset: Trained on a proprietary 200k-pair synthetic dataset luganda-synthetic-200k.
• Reproducibility: All benchmarks, significance tests, and scripts are hosted in this repository.

Engineering Pipeline

1. Architecture Optimization: Decoder pruned from 24 to 14 layers; vocabulary trimmed from 256k to 51k tokens to eliminate noise and reduce footprint.
2. Data Synthesis: 200,000 English pairs sourced from Wikinews/Wikivoyage with strict FLORES-200 overlap exclusion. Used MBR selection for the primary 100k set.
3. Knowledge Distillation: Retrained via word-level distillation (top-16 distribution + label-smoothed cross-entropy) on A100 hardware.
4. Optimization: Checkpoint averaging applied to the final 8 iterations for stability.
5. Inference: MBR decoding (beam-4 + 4 samples) utilized to maximize translation accuracy via chrF++ consensus.

Implementation

pip install ctranslate2 transformers sentencepiece huggingface_hub
import ctranslate2
from huggingface_hub import snapshot_download
from transformers import AutoTokenizer

path = snapshot_download("kyleparrratt/luganda-translator-v1")
tok = AutoTokenizer.from_pretrained("facebook/nllb-200-distilled-600M", src_lang="eng_Latn")
translator = ctranslate2.Translator(path, device="cpu", compute_type="int8")

def translate(text, beam_size=4):
    tokens = tok.convert_ids_to_tokens(tok.encode(text))
    result = translator.translate_batch([tokens], target_prefix=[["lug_Latn"]], beam_size=beam_size)
    return tok.decode(tok.convert_tokens_to_ids(result[0].hypotheses[0][1:]), skip_special_tokens=True)

print(translate("Where is the nearest hospital?"))
# Eddwaliro eriri okumpi liri ludda wa?