Loading…
Loading…
Model Description: This is a fine-tuned checkpoint of electra-small-discriminator, fine-tuned on an Australian University's student evaluation of teaching (SET) survey responses.
The model is used to automatically detect unacceptable comments left by students in SET surveys.
Examples of unacceptable comments are those containing abuse, insults, discrimination, indications of a risk of harm, or otherwise deemed unacceptable according to University policy.
Comments are labelled as Acceptable (0) or Unacceptable (1), with the probability of the classification used to a further discern which comments are screened manually by University staff.
This model serves as part of the Unversity's methodology for screening SET comments to prevent harm to the wellbeing and career prospects of academics.
The model can be used directly through our interactive Google Colab Notebook for ease of use.
Alternatively, it can be used through the transformers library. Here is an example:
import torch
from transformers import AutoTokenizer, ElectraForSequenceClassification
tokenizer = AutoTokenizer.from_pretrained("rickSomers/electra-small-QUT-unacceptable-student-feedback")
model = ElectraForSequenceClassification.from_pretrained("rickSomers/electra-small-QUT-unacceptable-student-feedback")
inputs = tokenizer("Thank you for your engaging tutorials", return_tensors="pt")
with torch.no_grad():
logits = model(**inputs).logits
predicted_class_id = logits.argmax().item()
model.config.id2label[predicted_class_id]
This model is intended be used for text classification. You can use this finetuned model directly for acceptable/unacceptable text classification, or it can be further fine-tune for downstream tasks.
This model should not be used without supporting processes and/or human evaluation. It should not be used to intentionally create hostile, harmful, or alienating environments for people.
Based on the nature of the training data being from a large, urban, Australian university, we observe that this model could produce biased predictions.
The performance of the model is therefore unknown in other contexts. While we believe that it could be beneficial for other instutes to adopt it in their SET screening, nuisanced differences in question type, language styles and choices, and local context could lead to varying performance.
We strongly advise users to thoroughly investigate their needs and use-case for this model, and evaluate its performance thoroughly before formally implementing into policies and processes. It is not recommended to use this model as the sole screening mechanism for harmful content.
The model was trained on 2021-2022 University SET qualitative survey responses. Each comment was asigned a label of 0 or 1, acceptable or unacceptable. Labels were asigned from previous keyword matching algorithms, deprecated unacceptable machine learning models, and staff raised lists.
Training was undertaken over several months to produce optimal performance. Oversampling techniques were used routinely to better balance the dataset due to a relatively small number of unacceptable comments.
The performance of the model on new SET responses are continually monitored and improved.
Internally, the model is routinely evaluated on the each semesters' data. To improve performance, it is often completely retrained from its base model again, or simply further fine-tuned with the inclusion of new text data.
Note that the model uploaded here is a deprecated version (trained on only 2021 & 2022 data) of the current model used by the university.