nllb-lua-en-mt-v1

This model is a bidirectional English (eng) ↔ Tshiluba (lua) translation model. It is a fine-tuned version of SalomonMetre13/nllb-lua-en-mt-v1, specifically optimized for translation in the Tshiluba language context.

Model Description

• Developed by: Salomon Metre
• Model Type: NLLB (No Language Left Behind) Encoder-Decoder
• Language(s): English (eng_Latn), Tshiluba (lua_Latn)
• License: CC-BY-NC-4.0
• Fine-tuned from: facebook/nllb-200-distilled-600M

Training and Evaluation Data

The model was fine-tuned on a parallel corpus of scraped Bible-based sentences. This dataset provides a critical foundation for Tshiluba, a low-resource language with limited digital parallel corpora.

• Dataset Link: https://huggingface.co/datasets/SalomonMetre13/lua_en
• Domain: Scriptural/Religious text.

Intended Uses & Limitations

Intended Use

• Research on machine translation for Congolese/Bantu languages.
• Practical drafting of translations between English and Tshiluba.

Limitations

• Domain Specificity: Performance is strongest on formal or scriptural text and may decrease on colloquial or highly technical English/Tshiluba.
• Morphological Complexity: As Tshiluba is a Bantu language with complex agglutinative morphology, the model may occasionally struggle with specific prefix/suffix agreements in out-of-distribution sentences.

Training Procedure

Training Hyperparameters

The following hyperparameters were used during training:

• Learning Rate: 3e-05
• Train Batch Size: 4
• Eval Batch Size: 4
• Optimizer: AdamW (Fused)
• LR Scheduler: Linear with 200 warmup steps
• Mixed Precision: Native AMP (FP16)

Evaluation Results (at step 4000)

• Eval Loss: 0.1439
• Epoch: 0.29 (partial epoch)

Framework Versions

• Transformers: 4.51.3
• Pytorch: 2.6.0+cu124
• Datasets: 3.6.0
• Tokenizers: 0.21.1