---
library_name: transformers
tags: []
---

# Model Card for MLotsawa

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

The model's name is taken from 'machine learning' (ML) and translation (Tibetan: Lotsawa).

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](https://pypi.org/project/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.

## 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](https://creativecommons.org/licenses/by-nc/4.0/)
- **Finetuned from model [optional]:** 'google-t5/t5-large'

### Model Sources [optional]

- **Repository:** [MLotsawa on Github](https://github.com/billingsmoore/MLotsawa)

## 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:

```python
from transformers import pipeline

translator = pipeline('translation', 'billingsmoore/mlotsawa')

input_text = <your transliterated Tibetan text>

translation = translator(input_text)

print(translation)
```

Note that if your input text is not already transliterated, you can transliterate unicode Tibetan into Wylie transliteration using the pyets library like so:

```python
import pyewts

converter = pyewts.pyewts()

input_text = <your unicode Tibetan text>

transliterated = converter.toWylie(input_text)
```

However, the model expects, and performs best on, Tibetan text that is transliterated in THL Simplified Phonetic Transcription rather than the Wylie 
transliterations that pyewts implements. A Python module to convert Wylie to THL is currently under developement. In the mean time, if you are familiar with Perl, 
[Roger Espel Llima has released a Perl module for this purpose here.](https://www.thlib.org/reference/transliteration/phconverter.php)

For now, the following script does reasonably well converting from the Wylie transliteration that is outputted by pyewts into something closer to phonetics.

```python

replacements = [[['lth', 'rh', 'db'] , ' '],
    [['rb', 'sb', 'sbr', 'lb', '’b', '\'b'] , 'b'],
    [['c', 'cw', 'gc', 'bc', 'lc', 'py', 'lpy', 'spy', 'dpy', 'mch', '’ch', '\'ch', 'phy', '’phy', '\'phy'] , 'ch'],
    [['rd', 'sd', 'gd', 'bd', 'brd', 'bsd', 'zl', 'bzl', 'ld', 'md', '’d', '\'d', 'dw'] , 'd'],
    [['rgr', 'lgr', 'sgr', 'dgr', 'dbr', 'bsgr', 'rbr', 'lbr', 'sbr', 'mgr', '’gr', '\'gr', '’dr', '\'dr', '’br', '\'br', 'gr', 'br', 'grw'] , 'dr'],
    [['rdz', 'gdz', 'brdz', 'mdz', '’dz', '\'dz'] , 'dz'],
    [['rg', 'lg', 'sg', 'dg', 'bg', 'brg', 'bsg', 'lg', 'mg', '’g', '\'g', 'gw'] , 'g'],
    [['rgy', 'lgy', 'sgy', 'dgy', 'bgy', 'brgy', 'bsgy', 'mgy', '’gy', '\'gy'] , 'gy'],
    [['hw'] , 'h'],
    [['rby', 'lby', 'sby', 'rj', 'gj', 'brj', 'lj', 'mj', '’j', '\'j', '’by', '\'by', 'by'] , 'j'],
    [['rk', 'lk', 'sk', 'kw', 'dk', 'bk', 'brk', 'bsk'] , 'k'],
    [['khw', 'mkh', '’kh', '\'kh'] , 'kh'],
    [['mkhy', '’khy', '\'khy'] , 'khy'],
    [['rky', 'lky', 'sky', 'dky', 'bky', 'brky', 'bsky'] , 'ky'],
    [['kl', 'gl', 'bl', 'rl', 'sl', 'brl', 'bsl', 'lw'] , 'l'],
    [['rm', 'sm', 'dm', 'smr', 'mr'] , 'm'],
    [['rn', 'sn', 'gn', 'brn', 'bsn', 'mn'] , 'n'],
    [['rng', 'lng', 'sng', 'dng', 'brng', 'bsng', 'mng'] , 'ng'],
    [['rny', 'sny', 'gny', 'brny', 'bsny', 'mny', 'nyw', 'rmy', 'smy', 'my'] , 'ny'],
    [['sp', 'dp', 'lp', 'ph', '’ph', '\'ph'] , 'p'],
    [['rw'] , 'r'],
    [['sr', 'sw', 'gs', 'bs', 'bsr'] , 's'],
    [['shw', 'gsh', 'bsh'] , 'sh'],
    [['rt', 'lt', 'st', 'tw', 'gt', 'bt', 'brt', 'blt', 'bst', 'bld', 'th', 'mth', '’th' '\'th'] , 't'],
    [['kr', 'rkr', 'lkr', 'skr', 'pr', 'lpr', 'spr', 'dkr', 'dpr', 'bkr', 'bskr', 'bsr', 'khr', 'thr', 'phr', 'mkhr', '’khr', '’phr'] , 'tr'],
    [['rts', 'sts', 'rtsw', 'stsw', 'gts', 'bts', 'brts', 'bsts', 'tsh', 'tshw', 'mtsh', '’tsh', '\'tsh'] , 'ts'],
    [['db', 'b'] , 'w'],
    [['g.y', 'dby'] , 'y'],
    [['zw', 'gz', 'bz'] , 'z'],
    [['zh', 'zhw', 'gzh', 'bzh'] , 'zh']]

def wylie_to_phonetic(wylie):
    phonetic = []
    for line in wylie:
        if line != '':
            # perform basic replacements
            result = line
            for elt in replacements:
                replace_list = elt[0]
                for string in replace_list:
                    result = result.replace(string, elt[1])

            # remove non-alphabetical chars
            result = re.sub(r'[^a-zA-Z\s]', '', result)
            phonetic.append(result)
    return phonetic

phonetic = wylie_to_phonetic(wylie)

```

### Downstream Use 

The model can be further finetuned using the following code:

```python
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/mlotsawa"
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.](https://www.kaggle.com/datasets/billingsmoore/classical-tibetan-to-english-translation-dataset)

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](https://pypi.org/project/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.