---
license: apache-2.0
datasets:
- NepaliAI/Nepali-HealthChat
- NepaliAI/Nepali-Health-Fact
language:
- en
- ne
metrics:
- bleu
pipeline_tag: text2text-generation
tags:
- health
- medical
- nlp
---

MT5-small is finetuned with large corups of Nepali Health Question-Answering Dataset.

### Training Procedure

The model was trained for 30 epochs with the following training parameters:

- Learning Rate: 2e-4
- Batch Size: 2
- Gradient Accumulation Steps: 8
- FP16 (mixed-precision training): Disabled
- Optimizer: AdamW with weight decay

The training loss consistently decreased, indicating successful learning.

## Use Case

```python

  !pip install transformers sentencepiece

  from transformers import MT5ForConditionalGeneration, AutoTokenizer 
  # Load the trained model
  model = MT5ForConditionalGeneration.from_pretrained("Chhabi/mt5-small-finetuned-Nepali-Health-50k-2")
  
  # Load the tokenizer for generating new output
  tokenizer = AutoTokenizer.from_pretrained("Chhabi/mt5-small-finetuned-Nepali-Health-50k-2",use_fast=True)


    
  query = "म धेरै थकित महसुस गर्छु र मेरो नाक बगिरहेको छ। साथै, मलाई घाँटी दुखेको छ र अलि टाउको दुखेको छ। मलाई के भइरहेको छ?"
  input_text = f"answer: {query}"
  inputs = tokenizer(input_text,return_tensors='pt',max_length=256,truncation=True).to("cuda")
  print(inputs)
  generated_text = model.generate(**inputs,max_length=512,min_length=256,length_penalty=3.0,num_beams=10,top_p=0.95,top_k=100,do_sample=True,temperature=0.7,num_return_sequences=3,no_repeat_ngram_size=4)
  print(generated_text)
  # generated_text
  generated_response = tokenizer.batch_decode(generated_text,skip_special_tokens=True)[0]
  tokens = generated_response.split(" ")
  filtered_tokens = [token for token in tokens if not token.startswith("<extra_id_")]
  print(' '.join(filtered_tokens))

```
## Evaluation
### BLEU score:


![image/png](https://cdn-uploads.huggingface.co/production/uploads/64a9a2e403835e13f9786936/X9NK63aj1EKeBH-d_NUG6.png)