finetuning-sentiment-model-bank_reviews-otherbank

This model is a fine-tuned version of distilbert/distilbert-base-uncased-finetuned-sst-2-english on app store reviews from OCBC bank and POSB bank (Singapore). It achieves the following results on the evaluation set:

• Loss: 0.4811
• Accuracy: 0.8630
• F1: 0.6970

Model description

Data was labelled according to review stars. If stars >3, review was ranked positive. Otherwise, it is labelled as negative. We have tried 4 stars instead of 3 as app developers would deem any negativity in reviews as negative as a whole, but accuracy dropped. Further investigations will need to be run. Above 4 stars positive: https://huggingface.co/ajiayi/finetuning-sentiment-model-bank_reviews-otherbank-4insteadof3 All data (OCBC,POSB,GXS): https://huggingface.co/ajiayi/finetuning-sentiment-model-bank_reviews

Intended uses & limitations

Model was used in the following project: https://github.com/weixuanontherun/DSA3101_Group-19 It was finetuned using OCBC and POSB and tested on GXS bank reviews. (GXS bank reviews NOT part of finetuning process)

Training and evaluation data

More information needed

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 2e-05
• train_batch_size: 16
• eval_batch_size: 16
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• num_epochs: 2

Training results

Model managed to achieve a 97-98% accuracy rate when run on GXS bank reviews.

Framework versions

• Transformers 4.38.2
• Pytorch 2.2.1+cu121
• Tokenizers 0.15.2