We all contrastive experience loss at some point in our lives, but which one for your problem and with what hyperparameters?

loss.py

# This will effectively train the embedding model. MultipleNegativesRankingLoss did not work.
loss: losses.ContrastiveLoss = losses.ContrastiveLoss(model=sbert_model)

# These are default arguments for OnlineContrastiveLoss
loss: losses.OnlineContrastiveLoss = losses.OnlineContrastiveLoss(
    model=sbert_model,
    margin=0.5,  # Margin for contrastive loss
    distance_metric=SiameseDistanceMetric.COSINE_DISTANCE,
)

loss: ContextAdaptiveContrastiveLoss = ContextAdaptiveContrastiveLoss(
    model=sbert_model,
    margin=0.5,  # Margin for contrastive loss
    gate_scale=5.0,  # Scale for the gate function
)