English-to-Embu-language-translation-LoRA

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

• Loss: 1.6197

Model description

A LoRA adapter for translating text from English to Kiembu.

Intended uses & limitations

More information needed

Training and evaluation data

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 0.0002
• train_batch_size: 8
• eval_batch_size: 8
• seed: 42
• gradient_accumulation_steps: 2
• total_train_batch_size: 16
• optimizer: Use OptimizerNames.ADAMW_TORCH_FUSED with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: linear
• num_epochs: 7
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss
4.8075	1.0	1741	2.1678
4.2583	2.0	3482	1.9215
3.9472	3.0	5223	1.7829
3.7003	4.0	6964	1.7045
3.6334	5.0	8705	1.6541
3.6319	6.0	10446	1.6279
3.5287	7.0	12187	1.6197

Evaluation metrics

Metric	Results	Explained
sacrebleu	10.895	Evaluation matches exact word sequence referencing our ground truth and adapter output.
		0-20= translation is literal or broken, 20-50=good, >60=very good or identical training and test data.
chrF++	41.718	Eval matches character sequence instead of whole words.
		Model adapter is getting the root words and grammar concepts right but using different word variations or spellings than our ground truth, this is concluded due to the low sacrebleu results.

Framework versions

• PEFT 0.19.1
• Transformers 5.0.0
• Pytorch 2.10.0+cu128
• Datasets 5.0.0
• Tokenizers 0.22.2