nllb-600M-lat_Latn-to-deu_Latn

NLLB-200-Distilled-600M - Latin to German Translation

This model is a fine-tuned version of the Meta NLLB-200-Distilled-600M model, specifically trained for Machine Translation (MT) from Latin (la) to German (de).

Model Details

Intended Use

This model is primarily designed for the direct translation of texts from the Latin language into the German language. It is suitable for academic, historical, and general translation tasks.

Limitations and Bias

• The model was trained on a specific dataset, and its coverage and style influence the translation quality.
• Like all large language models, it may contain errors, hallucinations, and bias stemming from the training data. Given that Latin is a language with limited modern corpora, translating it may present specific challenges with rare words, complex sentence structures, or contextual ambiguity.
• No commercial use is permitted, as it is released under the CC BY-NC 4.0 (NonCommercial) license.

Training Code

You can find the training-code on GitHub.

Training Data

You can find a Latin-German dataset also on Hugging Face.

📚 Citation / References

If you use this model in your research, please cite the underlying master's thesis as follows:

Masterarbeit (Zenodo DOI):

Wenzel, M. (2025). Translatio ex Machina: Neuronale Maschinelle Übersetzung vom Lateinischen ins Deutsche [Zenodo]. Unveröffentlichte Masterarbeit, Fachhochschule Südwestfalen

DOI: 10.5281/zenodo.17940090

How to Use the Model

You can use this model directly with the Hugging Face transformers library.

from transformers import AutoModelForSeq2SeqLM, AutoTokenizer

model_name = "fhswf/nllb-600M-lat_Latn-to-deu_Latn" 

tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSeq2SeqLM.from_pretrained(model_name)

SOURCE_LANGUAGE = "lat_Latn"
TARGET_LANGUAGE = "deu_Latn"

# Latin input text
latin_text = "Veni, vidi, vici." 

tokenizer.src_lang = SOURCE_LANGUAGE
inputs = tokenizer(latin_text, return_tensors="pt")

# Generating the translation
translated_tokens = model.generate(
    **inputs,
    max_length=128,
    num_beams=4,
    early_stopping=True
)

translated_tokens = model.generate(
    **inputs,
    forced_bos_token_id=tokenizer.convert_tokens_to_ids(TARGET_LANGUAGE),
    max_length=128,
    num_beams=4,
    early_stopping=True
)

# Decoding the result
translation = tokenizer.batch_decode(translated_tokens, skip_special_tokens=True)[0]

print(f"Latin: {latin_text}")
print(f"German: {translation}") 
# Expected Output: German: Ich kam, sah und siegte.