High Cost from Rollout Sampling — Is $6000+ Normal for 80k Prompts with 16 Rollouts Each?

[AlexFanw]

Hi team,

Thanks for the great work on verl.

I have a question regarding the cost of rollout sampling during training.

In my training setup:

Dataset size: 80,000 prompts
Each prompt generates 16 rollouts

Using an external API (e.g., Serper)
The total cost ends up being more than $6000, just for the rollout stage, mainly due to the large number of API queries (80,000 × 16 × n = 1.28n million).

I’d like to know the cost of your training?

Thank you!