marianMT_hin_eng_cs

This model is a fine-tuned version of Helsinki-NLP/opus-mt-mul-en on ar5entum/hindi-english-code-mixed dataset. It achieves the following results on the evaluation set:

• Loss: 0.1450
• Bleu: 77.8649
• Gen Len: 74.8945

Model description

The model is specifically designed to translate Hindi text written in Devanagari script into a mixed format where Hindi words are retained in Devanagari while English words are converted to Roman script. This model effectively handles the complexities of code-switching, producing output that accurately reflects the intended language mixing.

Example:

Hindi	Hindi + English CS
तो वो टोटली मेरे घर के प्लान पे डिपेंड करता है	to वो totally मेरे घर के plan पे depend करता है
मांग लो भाई बहुत नेसेसरी है	मांग लो भाई बहुत necessary है
टेलीविज़न में क्या प्रोग्राम चल रहा है?	television में क्या program चल रहा है?

from transformers import MarianMTModel, MarianTokenizer

class HinEngCS:
    def __init__(self, model_name='ar5entum/marianMT_hin_eng_cs'):
        self.model_name = model_name
        self.tokenizer = MarianTokenizer.from_pretrained(model_name)
        self.model = MarianMTModel.from_pretrained(model_name)

    def predict(self, input_text):
        tokenized_text = self.tokenizer(input_text, return_tensors='pt')
        translated = self.model.generate(**tokenized_text)
        translated_text = self.tokenizer.decode(translated[0], skip_special_tokens=True)
        return translated_text
model = HinEngCS()

input_text = "आज मैं नानयांग टेक्नोलॉजिकल यूनिवर्सिटी में अनेक समझौते होते हुए देखूंगा जो कि उच्च शिक्षा साइंस टेक्नोलॉजी और इनोवेशन में हमारे सहयोग को और बढ़ाएंगे।"
model.predict(input_text)
# आज मैं नानयांग technological university में अनेक समझौते होते हुए देखूंगा जो कि उच्च शिक्षा science technology और innovation में हमारे सहयोग को और बढ़ाएंगे।

Training Procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 5e-05
• train_batch_size: 50
• eval_batch_size: 50
• seed: 42
• distributed_type: multi-GPU
• num_devices: 2
• total_train_batch_size: 100
• total_eval_batch_size: 100
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• num_epochs: 30.0

Training results

Training Loss	Epoch	Step	Bleu	Gen Len	Validation Loss
1.5823	1.0	1118	11.6257	77.1622	1.1778
0.921	2.0	2236	33.2917	76.1459	0.6357
0.6472	3.0	3354	47.3533	75.9194	0.4504
0.5246	4.0	4472	55.2169	75.6871	0.3579
0.4228	5.0	5590	60.8262	75.5777	0.3041
0.3745	6.0	6708	64.8987	75.4424	0.2693
0.3552	7.0	7826	67.7607	75.2438	0.2455
0.3324	8.0	8944	69.635	75.1036	0.2274
0.2912	9.0	10062	71.3086	75.0326	0.2117
0.2591	10.0	11180	72.392	74.9607	0.2001
0.2471	11.0	12298	73.4758	74.9251	0.1899
0.236	12.0	13416	74.4219	74.833	0.1822
0.2265	13.0	14534	75.1435	74.9069	0.1745
0.2152	14.0	15652	75.7614	74.7409	0.1695
0.2078	15.0	16770	76.2353	74.7092	0.1641
0.2048	16.0	17888	76.7381	74.7274	0.1593
0.1975	17.0	19006	76.9954	74.7217	0.1559
0.1943	18.0	20124	77.421	74.6641	0.1524
0.1987	19.0	21242	77.8231	74.6833	0.1495
0.1855	20.0	22360	78.0784	74.6804	0.1472

Framework versions

• Transformers 4.45.0.dev0
• Pytorch 2.4.0+cu121
• Datasets 2.21.0
• Tokenizers 0.19.1