HingMaskedLM
MaskedLM is a pre-training technique used in Natural Language Processing (NLP) for deep-learning models like Transformers. It is a variant of language modeling where a portion of the input text is masked, and the model is trained to predict the masked tokens based on the context provided by the unmasked tokens. This model is trained for Masked Language Modeling for Hinglish Data.
Dataset
Hinglish-Top Dataset columns
- en_query
- cs_query
- en_parse
- cs_parse
- domain
Training
| Epoch | Loss |
|---|---|
| 1 | 0.0465 |
| 2 | 0.0262 |
| 3 | 0.0116 |
| 4 | 0.00385 |
| 5 | 0.0103 |
| 6 | 0.00738 |
| 7 | 0.00892 |
| 8 | 0.00379 |
| 9 | 0.00126 |
| 10 | 0.000684 |
Inference
from transformers import AutoTokenizer, AutoModelForMaskedLM, pipeline
tokenizer = AutoTokenizer.from_pretrained("SRDdev/HingMaskedLM")
model = AutoModelForMaskedLM.from_pretrained("SRDdev/HingMaskedLM")
fill = pipeline('fill-mask', model=model, tokenizer=tokenizer)
fill(f'please {fill.tokenizer.mask_token} ko cancel kardo')
Citation
Author: @SRDdev
Name: Shreyas Dixit
framework: Pytorch
Year: Jan 2023
Pipeline: fill-mask
Github: https://github.com/SRDdev
LinkedIn: https://www.linkedin.com/in/srddev/
