Fine-tune spoiler detection model

[ONDAAA]

Train a custom model on labeled spoiler data:

Data collection:

• Collect comments with explicit timestamps as weak labels
• Use GPT-4o as labeler on sample comments
• Human-in-the-loop validation from user feedback

Training:

• Base model: distilbert-base-multilingual-cased
• Task: regression (predict normalized timestamp 0-1)
• Alternative: classification into time buckets

Eval:

• ROC analysis for confidence threshold calibration
• A/B test vs embedding similarity baseline
• Target: >95% accuracy

Prerequisite: need ~5000+ labeled samples