longformer_4096_qsi

This model is a fine-tuned version of mrm8488/longformer-base-4096-finetuned-squadv2 on a tiny NovelQSI dataset. It achieves the following results on the evaluation set:

• Loss: 2.9598

Model description

This model is a test model for my research project. The idea of the model is to understand which novel character said the requested quote. It achieves a bit better results on the ´test´ split of the NovelQSI dataset than base longformer-base-4096-finetuned-squadv2 model on the same dataset split.

Base model results:

{
  "exact_match": {
    "confidence_interval": [8.754452551305853, 14.718614718614718],
    "score": 12.121212121212121,
    "standard_error": 1.8579217243778676
  },
  "f1": {
    "confidence_interval": [18.469101076147584, 28.28409063313956],
    "score": 22.799422799422796,
    "standard_error": 2.896728175757627
  },
  "latency_in_seconds": 0.7730605573419919,
  "samples_per_second": 1.2935597224598967,
  "total_time_in_seconds": 178.5769887460001
}

Achieved results:

{
  "exact_match": {
    "confidence_interval": [16.017316017316016, 24.242424242424242],
    "score": 20.346320346320347,
    "standard_error": 2.9434375492784994
  },
  "f1": {
    "confidence_interval": [23.123469058324783, 31.823648733317036],
    "score": 26.580086580086572,
    "standard_error": 2.593030474995015
  },
  "latency_in_seconds": 0.8093855569913422,
  "samples_per_second": 1.235505120349827,
  "total_time_in_seconds": 186.96806366500005
}

The results have shown, that the technique has its future.

Training and evaluation data

You can find training code in the github repo of my research:

https://github.com/Kkordik/NovelQSI

It was trained and evaluated in notebooks, so it is easy to reproduce.

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 2e-05
• train_batch_size: 1
• eval_batch_size: 1
• seed: 42
• gradient_accumulation_steps: 8
• total_train_batch_size: 8
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• num_epochs: 3

Training results

Framework versions

• Transformers 4.35.2
• Pytorch 2.1.0+cu118
• Datasets 2.15.0
• Tokenizers 0.15.0