Edit model card

This model can be used for Extractive QA

It has been finetuned for 3 epochs on SQuAD2.0.

Usage

from transformers import pipeline

qa_model = pipeline("question-answering", "timpal0l/mdeberta-v3-base-squad2")
question = "Where do I live?"
context = "My name is Tim and I live in Sweden."
qa_model(question = question, context = context)
# {'score': 0.975547730922699, 'start': 28, 'end': 36, 'answer': ' Sweden.'}

Evaluation on SQuAD2.0 dev set

{
    "epoch": 3.0,
    "eval_HasAns_exact": 79.65587044534414,
    "eval_HasAns_f1": 85.91387795001529,
    "eval_HasAns_total": 5928,
    "eval_NoAns_exact": 82.10260723296888,
    "eval_NoAns_f1": 82.10260723296888,
    "eval_NoAns_total": 5945,
    "eval_best_exact": 80.8809904826076,
    "eval_best_exact_thresh": 0.0,
    "eval_best_f1": 84.00551406448994,
    "eval_best_f1_thresh": 0.0,
    "eval_exact": 80.8809904826076,
    "eval_f1": 84.00551406449004,
    "eval_samples": 12508,
    "eval_total": 11873,
    "train_loss": 0.7729689576483615,
    "train_runtime": 9118.953,
    "train_samples": 134891,
    "train_samples_per_second": 44.377,
    "train_steps_per_second": 0.925
}

DeBERTaV3: Improving DeBERTa using ELECTRA-Style Pre-Training with Gradient-Disentangled Embedding Sharing

DeBERTa improves the BERT and RoBERTa models using disentangled attention and enhanced mask decoder. With those two improvements, DeBERTa out perform RoBERTa on a majority of NLU tasks with 80GB training data.

In DeBERTa V3, we further improved the efficiency of DeBERTa using ELECTRA-Style pre-training with Gradient Disentangled Embedding Sharing. Compared to DeBERTa, our V3 version significantly improves the model performance on downstream tasks. You can find more technique details about the new model from our paper.

Please check the official repository for more implementation details and updates.

mDeBERTa is multilingual version of DeBERTa which use the same structure as DeBERTa and was trained with CC100 multilingual data. The mDeBERTa V3 base model comes with 12 layers and a hidden size of 768. It has 86M backbone parameters with a vocabulary containing 250K tokens which introduces 190M parameters in the Embedding layer. This model was trained using the 2.5T CC100 data as XLM-R.

Downloads last month
134,924
Safetensors
Model size
278M params
Tensor type
I64
Β·
F32
Β·
Inference API

Model tree for timpal0l/mdeberta-v3-base-squad2

Finetuned
(206)
this model
Finetunes
14 models

Dataset used to train timpal0l/mdeberta-v3-base-squad2

Spaces using timpal0l/mdeberta-v3-base-squad2 22