AI-Ahmed/deberta-v3-base-funetuned-cls-qqa

A fine-tuned model based on the DeBERTaV3 model of Microsoft and fine-tuned on Glue QQP, which detects the linguistical similarities between two questions and whether they are duplicates questions or different.

Model Hyperparameters

epoch=4
per_device_train_batch_size=32
per_device_eval_batch_size=16
lr=2e-5
weight_decay=1e-2
gradient_checkpointing=True
gradient_accumulation_steps=8

Model Performance

{"Training Loss": 0.132400,
 "Validation Loss": 0.217410,
 "Validation Accuracy": 0.917969
}

Model Dependencies

{"Main Model": "microsoft/deberta-v3-base",
 "Dataset": "SetFit/qqp"
}

Training Monitoring & Performance

• wandb - deberta_qqa_classification

Model Testing

import torch
from transformers import AutoTokenizer, AutoModelForSequenceClassification
model_name = "AI-Ahmed/deberta-v3-base-funetuned-cls-qqa"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSequenceClassification.from_pretrained(model_name)
tokenized_input = tokenizer("How is the life of a math student? Could you describe your own experiences? Which level of preparation is enough for the exam jlpt5?", return_tensors="pt")

with torch.no_grad():
  logits = model(**tokenized_input).logits

predicted_class_id = logits.argmax().item()
model.config.id2label[predicted_class_id]

Information Citation

@inproceedings{
he2021deberta,
title={DEBERTA: DECODING-ENHANCED BERT WITH DISENTANGLED ATTENTION},
author={Pengcheng He and Xiaodong Liu and Jianfeng Gao and Weizhu Chen},
booktitle={International Conference on Learning Representations},
year={2021},
url={https://openreview.net/forum?id=XPZIaotutsD}
}