Model Card for phonetic-tibetan-to-english-translation

This model has been superseded by the model 'billingsmoore/tibetan-to-english-translation'

This model is a neural machine translation model for translating Literary Tibetan to English.

The model expects Tibetan transliterated according to THL Simplified Phonetic Transliteration as an input and outputs an English translation.

The model was evaluated using the BLEU metric as implemented by sacreBLEU, with a final score of 83.4374 on evaluation data. However, this score is unusually high, and may be the result of testing error. Stricter evaluation and training are currently in progress.

This work is licensed under Creative Commons Attribution-NonCommercial 4.0 International

Model Details

Model Description

This model is a finetuned T5 model with 770 million parameters.

• Developed by: billingsmoore
• Model type: [More Information Needed]
• Language(s) (NLP): Tibetan, English
• License: Attribution-NonCommercial 4.0 International
• Finetuned from model [optional]: 'google-t5/t5-large'

Model Sources [optional]

• Repository: MLotsawa on Github

Uses

This model is intended to be used as the translation model in the larger MLotsawa software, but can also be used in a Jupyter notebook or Python script.

Direct Use

To use this model for translation you can use the following code:

from transformers import pipeline

translator = pipeline('translation', 'billingsmoore/phonetic-tibetan-to-english-translation')

input_text = <your transliterated Tibetan text>

translation = translator(input_text)

print(translation)

Note that if your input text is not already transliterated in THL Simplified Phonetic Transliteration, you can transliterate unicode Tibetan using the thl-phonetic-transliteration module like so:

from thl_phonetic_transliteration.transliterator import Transliterator 

tibetan_text = '<your Tibetan unicode>'

transliterator = Transliterator()

thl_phonetics = transliterator.convert(tibetan_text)

The above module uses the model 'billingsmoore/tibetan-phonetic-transliteration'. More information about that model and how to use it can be found by clicking that link.

Downstream Use

The model can be further finetuned using the following code:

from datasets import load_dataset
from transformers import (
  AutoTokenizer, DataCollatorForSeq2Seq,
  AutoModelForSeq2SeqLM, Seq2SeqTrainingArguments,
  Seq2SeqTrainer, EarlyStoppingCallback, Adafactor
)
import evaluate
import numpy as np
from accelerate import Accelerator

data = load_dataset(<path_to_your_dataset>)

checkpoint = "billingsmoore/phonetic-tibetan-to-english-translation"
tokenizer = AutoTokenizer.from_pretrained(checkpoint)
data_collator = DataCollatorForSeq2Seq(tokenizer=tokenizer, model=checkpoint)

source_lang = 'bo'
target_lang = 'en'
prefix = "translate Tibetan to English: "

def preprocess_function(examples):

    inputs = [prefix + example[source_lang] for example in examples['translation']]
    targets = [example[target_lang] for example in examples['translation']]
    
    model_inputs = tokenizer(inputs, text_target=targets, max_length=128, truncation=True)

    return model_inputs

tokenized_dataset = dataset.map(preprocess_function, batched=True)

metric = evaluate.load("sacrebleu")

def postprocess_text(preds, labels):
    preds = [pred.strip() for pred in preds]
    labels = [[label.strip()] for label in labels]

    return preds, labels


def compute_metrics(eval_preds):
    preds, labels = eval_preds
    if isinstance(preds, tuple):
        preds = preds[0]
    decoded_preds = tokenizer.batch_decode(preds, skip_special_tokens=True)

    labels = np.where(labels != -100, labels, tokenizer.pad_token_id)
    decoded_labels = tokenizer.batch_decode(labels, skip_special_tokens=True)

    decoded_preds, decoded_labels = postprocess_text(decoded_preds, decoded_labels)

    result = metric.compute(predictions=decoded_preds, references=decoded_labels)
    result = {"bleu": result["score"]}

    prediction_lens = [np.count_nonzero(pred != tokenizer.pad_token_id) for pred in preds]
    result["gen_len"] = np.mean(prediction_lens)
    result = {k: round(v, 4) for k, v in result.items()}
    return result

early_stop = EarlyStoppingCallback()

model = AutoModelForSeq2SeqLM.from_pretrained(checkpoint, device_map="auto")

optimizer = Adafactor(
    model.parameters(), 
    scale_parameter=True, 
    relative_step=False, 
    warmup_init=False, 
    lr=3e-4
)

training_args = Seq2SeqTrainingArguments(
    output_dir=".",
    auto_find_batch_size=True,
    predict_with_generate=True,
    fp16=False, #check this
    push_to_hub=False,
    eval_strategy='epoch',
    save_strategy='epoch',
    load_best_model_at_end=True
)

trainer = Seq2SeqTrainer(
    model=model,
    args=training_args,
    train_dataset=tokenized_dataset['train'],
    eval_dataset=tokenized_dataset['test'],
    tokenizer=tokenizer,
    optimizers=(optimizer, None),
    data_collator=data_collator,
    compute_metrics=compute_metrics,
    callbacks=[early_stop]
)

trainer.train()

Training Details

Training Data

Training Data for this project is available here.

This dataset consists of 100,000 pairs of sentences or phrases. The first member of each pair is a sentence or phrase in Classical Tibetan. The second member is the English translation of the first.

The pairs are pulled from texts sourced from Lotsawa House (lotsawahouse.org) and are offered under the same license as the original texts they provided.

This data was scraped, cleaned, and formatted programmatically.

Training Procedure

This model was trained for 6 epochs on the dataset described above.

Training Hyperparameters

• This model was trained using the Adafactor optimizer with a learning rate of 2e-5.

Evaluation

The evaluation metric for this model was the BLEU score as implemented by sacreBLEU. BLEU (Bilingual Evaluation Understudy) scores measure the quality of machine-generated translations by comparing them to human-provided reference translations. The score ranges from 0 to 100, where 100 represents a perfect match with the reference translations. It evaluates the precision of n-grams (word sequences) in the generated text, with higher scores indicating closer alignment to the reference translations. A brevity penalty is applied to discourage translations that are too short.

The final BLEU score was 83.4374.