translategemma-4b-it-nb-nn

A fine-tuned Gemma 3 4B Instruct model for translating Norwegian Bokmål (nb) to Norwegian Nynorsk (nn) intended for deployment testing by the NB-ASR beta program.

Uploaded: 07-04-2026

The immediate purpose of this release is to support:

-reproducible beta evaluation,

-loading and inference validation in realistic environments,

-and packaging of a reviewed checkpoint for Hugging Face distribution.

Confidential beta release: this model card and the associated weights are intended for approved evaluators and collaborators. Treat the checkpoint as beta material rather than a public production release.

Model Description

This model was fine-tuned from google/gemma-3-4b-it on the NbAiLab/merged_npk_ndla_parallel_paragraphs dataset.

Intended Use

• Primary use: Translating Norwegian Bokmål text to Norwegian Nynorsk
• Language pair: nb → nn

Training Data

The model was trained on NbAiLab/merged_npk_ndla_parallel_paragraphs, a merged corpus of parallel Bokmål–Nynorsk paragraphs from NPK and NDLA.

Training Details

Parameter	Value
Base model	google/gemma-3-4b-it
Epochs	3
Global steps	46,728
Precision	bfloat16
Optimizer	AdamW (β1=0.9, β2=0.999, ε=1e-8)
Weight decay	—
Warmup ratio	0.1
Eval strategy	Every 2,000 steps
Dataloader workers	4
Train samples/sec	103.05
Train runtime	~8.1 hours

Evaluation Results

Evaluated on two test sets at the end of training (epoch 3):

NbAiLab Test Set (in-domain)

Metric	Score
BLEU	89.02
chrF	95.37
Loss	0.0750

Tatoeba nb→nn (out-of-domain)

Metric	Score
BLEU	72.20
chrF	85.68
Loss	0.4106

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "NbAiLab/translategemma-4b-it-nb-nn"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype="bfloat16", device_map="auto")

def translate_nb_to_nn(text: str) -> str:
    prompt = (
        "<start_of_turn>user\n"
        "You are a professional Norwegian (no) to Norwegian Nynorsk (nn) translator. "
        "Your goal is to accurately convey the meaning and nuances of the original Norwegian text while "
        "adhering to Norwegian Nynorsk grammar, vocabulary, and cultural sensitivities. Produce only the "
        "Norwegian Nynorsk translation, without any additional explanations or commentary. Please translate "
        "the following Norwegian text into Norwegian Nynorsk:\n\n\n"
        f"{text}<end_of_turn>\n"
        "<start_of_turn>model\n"
    )
    inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
    outputs = model.generate(**inputs, max_new_tokens=512)
    # Decode only the newly generated tokens (skip the prompt)
    generated = outputs[0][inputs["input_ids"].shape[-1]:]
    return tokenizer.decode(generated, skip_special_tokens=True).strip()

text = "Dette er en setning på bokmål som skal oversettes til nynorsk."
print(translate_nb_to_nn(text))

License

This model is subject to the Gemma Terms of Use.