Thanks for the great work! The SVTRv2 README reports results for synthetic training on MJ+ST, but the config/model/log are still marked as TODO. I tried reproducing it by modifying the released Union14M configs and replacing the training data with MJ train/val/test and ST, but the performance is still below the reported result. I also tried adjusting epoch_num and warmup according to the number of iterations, but this led to about a 4-point drop on U14 compared with the released setting. Could you please share the official SVTRv2 synthetic training config, or confirm the key settings such as epochs, warmup, batch size, learning rate, max_ratio, and whether any sampling/filtering was used?
Thanks for the great work! The SVTRv2 README reports results for synthetic training on MJ+ST, but the config/model/log are still marked as TODO. I tried reproducing it by modifying the released Union14M configs and replacing the training data with MJ train/val/test and ST, but the performance is still below the reported result. I also tried adjusting epoch_num and warmup according to the number of iterations, but this led to about a 4-point drop on U14 compared with the released setting. Could you please share the official SVTRv2 synthetic training config, or confirm the key settings such as epochs, warmup, batch size, learning rate, max_ratio, and whether any sampling/filtering was used?