metadata

license: apache-2.0
tags:
  - generated_from_keras_callback
model-index:
  - name: distilbert_ORO_screen
    results: []
widget:
  - text: >-
      The Paris Agreement target of limiting global surface warming to 1.5–2◦C
      compared to pre-industrial levels by 2100 will still heavily impact the
      ocean. While ambitious mitigation and adaptation are both needed, the
      ocean provides major opportunities for action to reduce climate change
      globally and its impacts on vital ecosystems and ecosystem services. A
      comprehensive and systematic assessment of 13 global- and local-scale,
      ocean-based measures was performed to help steer the development and
      implementation of technologies and actions toward a sustainable outcome.
      We show that (1) all measures have tradeoffs and multiple criteria must be
      used for a comprehensive assessment of their potential, (2) greatest
      benefit is derived by combining global and local solutions, some of which
      could be implemented or scaled-up immediately, (3) some measures are too
      uncertain to be recommended yet, (4) political consistency must be
      achieved through effective cross-scale governance mechanisms, (5)
      scientific effort must focus on effectiveness, co-benefits, disbenefits,
      and costs of poorly tested as well as new and emerging measures.

distilbert_ORO_screen

This model is a fine-tuned version of distilbert-base-uncased to classify whether an article's text (title + abstract + keywords) is relevant to an Ocean-related option to mitigate or adapt to climate change. Note that this model was fit on the full 'seen' dataset, therefore predictions will only be approximations of those reported in the article (in prep.).
It achieves the following results on the evaluation set:

Train Loss: 0.3204
Train Binary Accuracy: 0.8819
Validation Loss: 1.5870
Validation Binary Accuracy: 0.3519
Epoch: 3

Model description

This model predicts a label that indicated 'relevance' for an ocean related option as a value between 0 and 1. Therefore a number > 0.5 indicates it is more likely to be relevant.

Intended uses & limitations

This model is intended to be applied to article text (title + abstract + keywords) retrieved from citation indexed databases such as Web of Science or Scopus using a search query. This can be used to autonomously screen for relevant articles from a large volume of literature and can be used in analyses that provide a granular map of the distribution of relevant studies. NB for more specific research questions, this should be complemented with manual literature review methodology (e.g. see the Collaboration for Environmental Evidence Guidelines )

Training and evaluation data

For a description of the dataset, see the paper (in prep.)

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

optimizer: {'name': 'AdamW', 'learning_rate': 1e-05, 'decay': 0.0, 'beta_1': 0.9, 'beta_2': 0.999, 'epsilon': 1e-07, 'amsgrad': False, 'weight_decay': 0.0, 'exclude_from_weight_decay': None}
training_precision: float32

Training results

Train Loss	Train Binary Accuracy	Validation Loss	Validation Binary Accuracy	Epoch
0.5592	0.6278	1.2920	0.2050	0
0.4488	0.7952	1.3261	0.2612	1
0.3743	0.8281	1.7808	0.2439	2
0.3204	0.8819	1.5870	0.3519	3

Framework versions

Transformers 4.30.2
TensorFlow 2.11.0
Datasets 2.18.0
Tokenizers 0.13.3