nllb-200-coreml-256

NLLB-200 CoreML (256 tokens)

On-device neural machine translation for 200 languages using CoreML on Apple devices (iPhone, iPad, Mac).

This is a CoreML conversion of facebook/nllb-200-distilled-600M optimized for:

• ✅ Fast on-device inference
• ✅ GPU/Neural Engine acceleration
• ✅ 256-token context (≈150-180 words)
• ✅ 2X longer context vs 128-token version

📦 What's Included

.
├── NLLB_Encoder_256.mlpackage    # Encoder model (~1.5 GB)
├── NLLB_Decoder_256.mlpackage    # Decoder model (~1.7 GB)
├── tokenizer/                     # Tokenizer files
├── example.py                     # Ready-to-run example
└── language_codes.json            # Language code reference

🚀 Quick Start

Installation

pip install coremltools transformers

Download Models

# Clone this repo
git lfs install
git clone https://huggingface.co/cstr/nllb-200-coreml-256
cd nllb-200-coreml-256

Run Translation

from example import translate_text

# English to German
result = translate_text(
    "Hello, how are you today?",
    source_lang="eng_Latn",
    target_lang="deu_Latn"
)
print(result)  # "Hallo, wie geht es dir heute?"

💡 Usage Examples

Multiple Languages

from example import translate_text

# English → Spanish
translate_text("Good morning!", "eng_Latn", "spa_Latn")
# → "¡Buenos días!"

# French → English
translate_text("Bonjour le monde", "fra_Latn", "eng_Latn")
# → "Hello world"

# Japanese → English
translate_text("こんにちは", "jpn_Jpan", "eng_Latn")
# → "Hello"

Long Text Translation

# 256-token context handles longer paragraphs
long_text = """
Machine learning is a subset of artificial intelligence that 
enables computers to learn and improve from experience without 
being explicitly programmed. In recent years, it has transformed 
technology and created new possibilities.
"""

result = translate_text(long_text, "eng_Latn", "deu_Latn")

Production Usage

import coremltools as ct
from transformers import AutoTokenizer

class Translator:
    def __init__(self):
        # Load once, reuse for all translations
        self.encoder = ct.models.MLModel(
            "NLLB_Encoder_256.mlpackage",
            compute_units=ct.ComputeUnit.ALL  # Use GPU
        )
        self.decoder = ct.models.MLModel(
            "NLLB_Decoder_256.mlpackage",
            compute_units=ct.ComputeUnit.ALL
        )
        self.tokenizer = AutoTokenizer.from_pretrained("./tokenizer")
        
    def translate(self, text, src_lang, tgt_lang):
        # Your translation logic here
        pass

# Create once
translator = Translator()

# Reuse many times (fast!)
translator.translate("Hello", "eng_Latn", "deu_Latn")
translator.translate("Goodbye", "eng_Latn", "fra_Latn")

🌍 Supported Languages

See language_codes.json for the full list of 200+ languages. Common examples:

Full list: NLLB Language Codes

⚙️ Technical Details

• Max Tokens: 256 (≈150-180 words depending on language)
• Precision: FLOAT16
• Compute: CPU + GPU + Neural Engine
• Base Model: facebook/nllb-200-distilled-600M
• Model Size: ~3.2 GB total (encoder + decoder)

🔧 Advanced Options

CPU-Only Mode

encoder = ct.models.MLModel(
    "NLLB_Encoder_256.mlpackage",
    compute_units=ct.ComputeUnit.CPU_ONLY
)

Batch Processing

texts = ["Hello", "Goodbye", "Thank you"]
translations = [translate_text(t, "eng_Latn", "deu_Latn") for t in texts]

📊 Comparison with 128-Token Version

⚠️ Limitations

• 256 token limit: Longer text is truncated (~150-180 words)
• Quality: Distilled model, slightly lower quality than full NLLB-3.3B
• Low-resource languages: May have reduced accuracy
• No streaming: Complete sentence processing only

📝 License

• Models: CC-BY-NC-4.0 (inherited from NLLB-200)
• Code: MIT

⚠️ Non-commercial use only per NLLB license

🔗 Related Models

• 128-token version - Faster for short texts

## Provenance and EU AI Act Art. 53 note

- **Upstream model:** [facebook/nllb-200-distilled-600M](https://huggingface.co/facebook/nllb-200-distilled-600M) — published by `facebook`.
- **Upstream licence:** `cc-by-nc-4.0`. This repository redistributes under the same terms; it grants no rights the upstream licence does not.
- **What was done here:** format conversion and/or quantisation only (CoreML). No training, no fine-tuning, no merging, no distillation, no change to architecture, vocabulary or capability. Only the numeric representation of the upstream weights differs.
- **Training data:** documented — where it is documented at all — by the upstream provider; see the upstream model card. No training data was used, added or selected by this repository.
- **Provider status:** under Regulation (EU) 2024/1689 the upstream authors remain the provider of this model. Converting the serialisation format does not make this repository the provider of a new general-purpose AI model, and no such claim is made. Questions about training content, copyright policy or model capability belong upstream.