<!-- This model card has been generated automatically according to the information the Trainer had access to. You should probably proofread and complete it, then remove this comment. -->

Important Note:

I created the combined metric (55% F1 score + 45% exact match score) and load the state with the best result at the end. Here is the setting in the TrainingArguments:

  load_best_model_at_end=True,
  metric_for_best_model='combined',
  greater_is_better=True,

DSPFirst-Finetuning-5

This model is a fine-tuned version of ahotrod/electra_large_discriminator_squad2_512 on a generated Questions and Answers dataset from the DSPFirst textbook based on the SQuAD 2.0 format.<br /> It achieves the following results on the evaluation set:

• Loss: 0.8529
• Exact: 67.0964
• F1: 74.4842
• Combined: 71.1597

More accurate metrics:

Before fine-tuning:

 'HasAns_exact': 54.71817606079797,
 'HasAns_f1': 61.08672724332754,
 'HasAns_total': 1579,
 'NoAns_exact': 88.78048780487805,
 'NoAns_f1': 88.78048780487805,
 'NoAns_total': 205,
 'best_exact': 58.63228699551569,
 'best_exact_thresh': 0.0,
 'best_f1': 64.26902596256402,
 'best_f1_thresh': 0.0,
 'exact': 58.63228699551569,
 'f1': 64.26902596256404,
 'total': 1784

After fine-tuning:

 'HasAns_exact': 67.57441418619379,
 'HasAns_f1': 75.92137683558988,
 'HasAns_total': 1579,
 'NoAns_exact': 63.41463414634146,
 'NoAns_f1': 63.41463414634146,
 'NoAns_total': 205,
 'best_exact': 67.0964125560538,
 'best_exact_thresh': 0.0,
 'best_f1': 74.48422310728503,
 'best_f1_thresh': 0.0,
 'exact': 67.0964125560538,
 'f1': 74.48422310728503,
 'total': 1784

Dataset

A visualization of the dataset can be found here.<br /> The split between train and test is 70% and 30% respectively.

DatasetDict({
    train: Dataset({
        features: ['id', 'title', 'context', 'question', 'answers'],
        num_rows: 4160
    })
    test: Dataset({
        features: ['id', 'title', 'context', 'question', 'answers'],
        num_rows: 1784
    })
})

Intended uses & limitations

This model is fine-tuned to answer questions from the DSPFirst textbook. I'm not really sure what I am doing so you should review before using it.<br /> Also, you should improve the Dataset either by using a better generated questions and answers model (currently using https://github.com/patil-suraj/question_generation) or perform data augmentation to increase dataset size.

Training and evaluation data

• batch_size of 6 results in 14.03 GB VRAM
• Utilizes gradient_accumulation_steps to get total batch size to 516 (total batch size should be at least 256)
• 4.52 GB RAM
• 30% of the total questions is dedicated for evaluating.

Training procedure

• The model was trained from Google Colab
• Utilizes Tesla P100 16GB, took 6.3 hours to train
• load_best_model_at_end is enabled in TrainingArguments

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 2e-05
• train_batch_size: 6
• eval_batch_size: 6
• seed: 42
• gradient_accumulation_steps: 86
• total_train_batch_size: 516
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• num_epochs: 7

Model hyperparameters

• hidden_dropout_prob: 0.36
• attention_probs_dropout_prob = 0.36

Training results

Framework versions

• Transformers 4.18.0
• Pytorch 1.10.0+cu111
• Datasets 2.1.0
• Tokenizers 0.12.1