VanModers114/East_Frisian_NLLB_Model

East Frisian NLLB Model

Model summary

VanModers114/East_Frisian_NLLB_Model is the main Oostfräisk Ooversetter model. It is a fine-tuned version of facebook/nllb-200-distilled-600M for translation between East Frisian Low Saxon, German, and English.

Supported and trained directions are:

• German (deu_Latn) → East Frisian (frs_Latn)
• East Frisian (frs_Latn) → German (deu_Latn)
• English (eng_Latn) → East Frisian (frs_Latn)
• East Frisian (frs_Latn) → English (eng_Latn)

This is a research-oriented low-resource translation model. It is not a certified translation service and should not be relied upon without human review for consequential decisions.

Model details

Property	Value
Developer	Oostfräisk Instituut, VanModers114
Model type	Encoder-decoder Transformer (M2M100ForConditionalGeneration)
Base model	facebook/nllb-200-distilled-600M
Parameters	NLLB distilled 600M class
Context limit	1,024 positions; fine-tuning truncates to 512 tokens
East Frisian language token	frs_Latn
License	CC BY-NC 4.0, inherited from the NLLB base model
Source code	VanModers/oostfraeisk_ooversetter

NLLB does not include a dedicated East Frisian token. Fine-tuning adds frs_Latn in an unused embedding slot and initializes that embedding from Dutch nld_Latn. German and English retain their native NLLB tokens.

Evaluation

The final model from the recorded run was evaluated German → East Frisian on a separate 100-pair project validation corpus:

Metric	Result
SacreBLEU	62.06
Teacher-forced token accuracy	0.8937
Average loss	0.8808

Important qualifications:

• Only German → East Frisian was evaluated in this run.
• The 100-pair set is small and is not a broad benchmark of dialects, domains, or sentence lengths.
• The SacreBLEU signature and confidence interval were not recorded.
• Token accuracy is measured under teacher forcing; it is not sentence accuracy and should not be interpreted as the probability that a generated translation is correct.

How to use

The model uses NLLB-style source and target language control. The compatibility setup below also handles the strict boolean/integer configuration validation present in transformers 5.4.0:

import json
import torch
from huggingface_hub import hf_hub_download
from transformers import AutoModelForSeq2SeqLM, AutoTokenizer, M2M100Config

MODEL_ID = "VanModers114/East_Frisian_NLLB_Model"

# Compatibility for strict transformers 5.4.0 validation of scale_embedding.
scale_field = getattr(M2M100Config, "__dataclass_fields__", {}).get("scale_embedding")
if scale_field is not None and isinstance(scale_field.default, bool):
    scale_field.default = int(scale_field.default)

config_path = hf_hub_download(repo_id=MODEL_ID, filename="config.json")
with open(config_path, encoding="utf-8") as config_file:
    config_dict = json.load(config_file)
if isinstance(config_dict.get("scale_embedding"), bool):
    config_dict["scale_embedding"] = int(config_dict["scale_embedding"])
config = M2M100Config.from_dict(config_dict)

tokenizer = AutoTokenizer.from_pretrained(MODEL_ID, config=config)
model = AutoModelForSeq2SeqLM.from_pretrained(MODEL_ID, config=config)
model.eval()

# Supported codes: deu_Latn, eng_Latn, frs_Latn
source_language = "deu_Latn"
target_language = "frs_Latn"
text = "Das ist ein Beispielsatz."

frs_id = tokenizer.convert_tokens_to_ids("frs_Latn")
if frs_id == tokenizer.unk_token_id:
    raise RuntimeError("The tokenizer does not contain the frs_Latn token")

# Compatibility for tokenizer versions that keep explicit language-code maps.
if hasattr(tokenizer, "lang_code_to_id"):
    tokenizer.lang_code_to_id["frs_Latn"] = frs_id
if hasattr(tokenizer, "id_to_lang_code"):
    tokenizer.id_to_lang_code[frs_id] = "frs_Latn"

tokenizer.src_lang = source_language
inputs = tokenizer(text, return_tensors="pt", truncation=True, max_length=512)
target_id = tokenizer.convert_tokens_to_ids(target_language)

with torch.inference_mode():
    output = model.generate(
        **inputs,
        forced_bos_token_id=target_id,
        max_new_tokens=512,
        num_beams=4,
    )

print(tokenizer.decode(output[0], skip_special_tokens=True))

The source repository's nllb_model/app.py contains the maintained Gradio implementation of the same loading approach.

License

This model is distributed under CC BY-NC 4.0, inherited from facebook/nllb-200-distilled-600M. Review the base model card and full license before use.

Citation

If you use the underlying NLLB architecture or weights, cite:

@article{nllb2022,
  title = {No Language Left Behind: Scaling Human-Centered Machine Translation},
  author = {{NLLB Team} and Costa-jussà, Marta R. and others},
  journal = {arXiv preprint arXiv:2207.04672},
  year = {2022}
}

Project model and training code: https://github.com/VanModers/oostfraeisk_ooversetter

Contact

Open an issue in the source repository or contact the model publisher through Hugging Face or by contacting the Oostfräisk Instituut.