This model is a fine-tuned version of facebook/nllb-200-distilled-600M on the NjeriKahoro/thiomi-multilingual-text dataset. It achieves the following results on the evaluation set:

Loss: 1.5709
Chrf: 46.2552
Bleu: 20.0191

Model description

This model is a sequence-to-sequence bidirectional translator tailored for high-fidelity conversion between Kikuyu (Gĩkũyũ) and English in both directions. By utilizing a symmetric training design, a single instance of this model handles text processing natively without requiring separate checkpoints for each language pair direction.

Intended Uses

Machine translation research for low-resource and morphologically rich African languages.
Symmetrical conversion between English and Kikuyu for localized digital content or chat assistants.

Limitations

Domain Focus: Performance is strongly aligned with the linguistic domain of the NjeriKahoro/thiomi-multilingual-text dataset. Performance may vary on highly specialized legal, financial, or ancient idiomatic texts.
Morphological Agglutination: As a low-resource model, it may occasionally hallucinate or mistranslate complex, highly agglutinative Kikuyu verb structures not widely captured during training.

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

learning_rate: 2e-05
train_batch_size: 2
eval_batch_size: 2
seed: 42
gradient_accumulation_steps: 16
total_train_batch_size: 32
optimizer: Use OptimizerNames.PAGED_ADAMW_8BIT with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
lr_scheduler_type: linear
num_epochs: 3
mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Chrf	Bleu
58.2375	1.0	341	1.6319	45.1751	18.9449
52.3836	2.0	682	1.5796	45.9867	19.8886
50.1786	3.0	1023	1.5709	46.2552	20.0191

Framework versions

Transformers 5.8.1
Pytorch 2.10.0+cu128
Datasets 4.8.5
Tokenizers 0.22.2

How to use

Because this model is built on NLLB-200, you must explicitly declare the standard BCP-47 language codes (kik_Latn for Kikuyu and eng_Latn for English) for proper target token routing.

Option 1: High-Level `pipeline` API (Recommended)

python

from transformers import pipeline

model_id = "NjeriKahoro/nllb-200-Thiomi-Kik-Eng"

# English to Kikuyu
en_to_ki = pipeline("translation", model=model_id, src_lang="eng_Latn", tgt_lang="kik_Latn", device=0)
print(en_to_ki("Hello, how are you?")[0]['translation_text'])

# Kikuyu to English
ki_to_en = pipeline("translation", model=model_id, src_lang="kik_Latn", tgt_lang="eng_Latn", device=0)
print(ki_to_en("Wĩ mwega?")[0]['translation_text'])

NjeriKahoro/nllb-200-Thiomi-Kik-Eng

nllb-200-Thiomi-Kik-Eng