Qartify NLLB Menu Translation

A fine-tuned NLLB-200 Distilled 600M model for multilingual restaurant menu translation, developed for the Qartify platform.

This repository contains:

the original fine-tuned PyTorch checkpoint
an optimized ONNX INT8 version for production inference
the dish-name protection lexicon used by the translation pipeline

Model Overview

Base Model

facebook/nllb-200-distilled-600M

Supported Languages

Arabic (arb_Arab)
French (fra_Latn)
English (eng_Latn)

Supported translation directions:

Arabic → French
Arabic → English
French → Arabic
French → English
English → Arabic
English → French

Training

The model was fine-tuned on a curated multilingual restaurant menu dataset containing:

restaurant menu items
dish names
food descriptions
multilingual parallel translations
manually reviewed and cleaned samples

The objective was to improve translation quality for restaurant menus while preserving culturally specific food names.

Evaluation

Translation Quality

Metric	Score
BLEU	49.77

The fine-tuned checkpoint significantly improves restaurant menu translation compared with the original NLLB model.

Dish Name Preservation

Dish-name preservation is achieved using a detect-and-substitute post-processing pipeline together with the provided dish lexicon.

Result

100% dish-name preservation on the protected evaluation dataset.

The translation model generates the translated sentence, while the post-processing pipeline restores protected dish names exactly.

Example

Arabic

طاجين دجاج بالخضر

↓

French

Tajine de poulet aux légumes

Arabic

كسكس عضوي

↓

English

Organic Couscous

Loading the PyTorch Checkpoint

from transformers import AutoTokenizer, AutoModelForSeq2SeqLM

tokenizer = AutoTokenizer.from_pretrained(
    "BlockwardSARL/qartify-nllb-menu-translation/checkpoint"
)

model = AutoModelForSeq2SeqLM.from_pretrained(
    "BlockwardSARL/qartify-nllb-menu-translation/checkpoint"
)

Example inference:

tokenizer.src_lang = "arb_Arab"

inputs = tokenizer(
    "طاجين دجاج بالخضر",
    return_tensors="pt"
)

generated = model.generate(
    **inputs,
    forced_bos_token_id=tokenizer.convert_tokens_to_ids("fra_Latn")
)

translation = tokenizer.batch_decode(
    generated,
    skip_special_tokens=True
)

print(translation[0])

ONNX Inference

This repository also provides an optimized ONNX INT8 Dynamic Quantized version intended for fast production inference with ONNX Runtime.

The ONNX model offers:

lower memory usage
faster inference
production-ready deployment

Dish Protection

The repository includes:

lexicon/
└── dish_lexicon.json

This lexicon is used by the translation service to preserve protected dish names through the detect-and-substitute pipeline.

Repository Structure

checkpoint/
├── config.json
├── model.safetensors
├── tokenizer.json
├── tokenizer_config.json
├── sentencepiece.bpe.model
├── generation_config.json
└── special_tokens_map.json

onnx_int8_dynamic/
├── encoder_model.onnx
├── decoder_model.onnx
├── decoder_with_past_model.onnx
└── ...

lexicon/
└── dish_lexicon.json

Intended Use

Recommended for:

restaurant menu translation
OCR menu pipelines
food ordering platforms
hospitality applications
multilingual restaurant management systems

Limitations

The model was fine-tuned specifically for restaurant menu translation.

Performance on general-domain text has not been evaluated.

Perfect dish-name preservation requires the accompanying dish lexicon and detect-and-substitute pipeline.

Citation

If you use this model in academic work, please cite:

Qartify NLLB Menu Translation
Blockward SARL
2026

Authors

Developed by Blockward SARL

Master Research Internship – Qartify AI

Ranya El Hamdaoui

Downloads last month: -; Downloads are not tracked for this model. How to track