ucsinfer/TODO.md

Recommend

• Use History when adding catids

Gather

• Maybe more dataset configurations

Qualify

• Print stats for a dataset

Fine-tune

• https://www.sbert.net/docs/sentence_transformer/loss_overview.html#loss-table
  • Use (anchor, positive) pairs to train a new model
  • Use (sentence) + class labels to train a new model
• Implement BatchAllTripletLoss
• Implement a two-phase training regime
  • 1. Train with anchored definitions then...
  • 2. Train with class labels

Evaluate

• Print more information about the dataset coverage of UCS

  • Allow skipping model testing for this

• Print raw output

Utility

• Clear caches