single mention · post-training
Reward-scale calibration
Reward-scale calibration designs task rewards to have broadly comparable scales, so heterogeneous reward signals can be combined across tasks.1,2 In the described design, reward values typically fall in [0,1], with bonuses and penalties able to shift the final rewards outside this range.3
How this definition was written
Summarised by glm-5.3-flash from 1 source; every sentence is checked against the quote it cites.
we design task rewards to have broadly comparable scales
— technical reportEach data source defines its own reward function by combining signals into a composite reward
— technical reportwith values typically in [0,1] . Bonuses and penalties can shift the final rewards outside this range
— technical report
Filed under post-training › reward modelling