writ-tech / Fine-Tuning-BART

SynthUse

Datasets and models used:

BART FINE TUNING

• Mediasum Dataset from huggingface.("ccdv/mediasum")
• BART model used from hugging face.( model_name = "sshleifer/distilbart-xsum-12-3")
• Tokenization process:

• Results before fine tuning

• Results after fine tuning

• Dataset size :
  1. Train= 5000 datapoints
  2. Validation =22 datapoints
  3. Test =22 datapoints
• Training Parameteres :

1. Batch Size=4
2. No of Epochs=1
3. Weight Deacy=0.1
4. label_smoothing_factor=0.1

• A snapshot from the training preiod.

A snapshot from the streamlit local server app.