madlad400-finetuned-kyu-eng

This model is a fine-tuned version of jbochi/madlad400-3b-mt for translation from Western Kayah to English.

Model details

• Developed by: SIL Global
• Finetuned from model: jbochi/madlad400-3b-mt
• Model type: Translation
• Source language: Western Kayah (kyu)
• Target language: English (eng)
• License: closed/private

Datasets

The model was trained on a parallel corpus of plain text files:

Western Kayah:

• The New Testament and Old Testament portions in the Kayah Li Language
• License: © Wycliffe Bible Translators

English:

• English Berean Standard Bible
• License: Inc. All rights reserved. Used by permission.

Usage

You can use this model with the transformers library like this:

from transformers import AutoTokenizer, AutoModelForSeq2SeqLM

tokenizer = AutoTokenizer.from_pretrained("sil-ai/madlad400-finetuned-kyu-eng")
model = AutoModelForSeq2SeqLM.from_pretrained("sil-ai/madlad400-finetuned-kyu-eng")

inputs = tokenizer("Your input text here", return_tensors="pt")
outputs = model.generate(**inputs)
print(tokenizer.decode(outputs[0]))

Framework versions

• PEFT 0.12.0
• Transformers 4.44.2
• Pytorch 2.4.1+cu124
• Datasets 2.21.0
• Tokenizers 0.19.1

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 0.0005
• train_batch_size: 4
• eval_batch_size: 32
• seed: 42
• gradient_accumulation_steps: 8
• total_train_batch_size: 32
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_ratio: 0.1
• num_epochs: 10.0

Training results

Note: Due to the limited size of the dataset, the same data was used for both training and evaluation, instead of using a separate holdout set for testing. These results may not be representative of the model's real-world performance.

Training Loss	Epoch	Step	Validation Loss	Chrf
0.7923	2.1415	1600	0.7236	34.6362
0.4648	4.2831	3200	0.3701	53.0197
0.3151	6.4246	4800	0.2263	62.5914
0.2346	8.5662	6400	0.1422	70.0758
		Final:	0.1216	72.7846