epoch setting for the CC-Finetune text supervised contrastive training stage

[Gbone3176]

Hello, this is truly a wonderful and inspiring work. During the reproduction process, I am a bit confused about the epoch setting of 1 for the CC-Finetune pure text fine-tuning stage. Why is only 1 epoch sufficient to learn enough text (mainly similar subtitle text here) representation ability? related configs here

Looking forward to your reply. Thanks again.

[raytrun]

Thanks! In our experiments, 1 epoch is already enough for convergence in the pure text finetuning stage. We didn’t see clear benefits from more epochs, so we chose to allocate the compute to more data instead.

[Gbone3176]

Thank you for your response. Are you referring to the benefits you mentioned as the direct results on the downstream CLIP task, or are you referring to the loss or other metrics during the Stage 1 training phase? I don't know much about representation learning, so I'm not very clear on these things. Thank you very much for your reply!