File size: 2,153 Bytes
fe11ff5 528ef20 fe11ff5 d807ce7 528ef20 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 |
---
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:
 |