Temperature Ablation

Key Insight

The temperature τ divides the cosine-similarity scores before the softmax in InfoNCE, tuning how sharp or soft that contest between the true pair and its hardest negatives becomes — and this project sweeps it end to end instead of trusting the default. Running τ from 0.01 to 1.0 and watching both accuracy and the geometry of the embedding space — how tightly true pairs cluster on the unit sphere — makes visible why CLIP makes τ a learned parameter (initialized around 0.07) rather than a guessed constant. Set it too low and training is unstable and overconfident; too high and the model never commits — which is why this one scalar punches far above its weight.