Clarification on GPU count used in reward model training

[LAOS-Y]

Are all the reward models trained with 32 GPUs in total, including models with single machine config, e.g., editscore_7B, editscore_qwen3_vl_4B_instruct, and editscore_qwen3_vl_8B_instruct?

Is all multi-machine reward training done with 2 nodes with 16 GPUs each?

[sabulin]

The released configs are standardized for reproducibility and may not exactly match every original run.

As far as I recall, the 7B-scale reward models were trained on a single machine. For larger models such as 32B/72B, we tried both 2-node and 4-node setups and did not observe noticeable performance differences.