THUDM

/

glm-10b-chinese

like 120

Follow

Knowledge Engineering Group (KEG) & Data Mining at Tsinghua University 1,289

The loss implementation of condition generation model is wrong. It takes me 3 days to debug!!!
The calculation for loss should be modified as: loss = loss_fct(lm_logits[:, :-1, :].reshape(-1, lm_logits.size(-1)), labels[:, 1:].reshape(-1))