From 4a98f4cedc8d7e3fbfff88bc4ef2bf5807cdd157 Mon Sep 17 00:00:00 2001 From: Mr-Neutr0n <64578610+Mr-Neutr0n@users.noreply.github.com> Date: Wed, 11 Feb 2026 23:46:59 +0530 Subject: [PATCH] fix: use eval_dataloader instead of train_dataloader in KTO evaluation --- applications/ColossalChat/coati/trainer/kto.py | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/applications/ColossalChat/coati/trainer/kto.py b/applications/ColossalChat/coati/trainer/kto.py index f87bf53c40cf..9a2322a050d4 100755 --- a/applications/ColossalChat/coati/trainer/kto.py +++ b/applications/ColossalChat/coati/trainer/kto.py @@ -273,11 +273,11 @@ def _eval(self, epoch: int): self.model.eval() self.accumulative_meter.reset() step_bar = trange( - len(self.train_dataloader) // self.accumulation_steps, + len(self.eval_dataloader) // self.accumulation_steps, desc=f"Epoch {epoch + 1}/{self.max_epochs}", disable=not is_rank_0(), ) - for i, batch in enumerate(self.train_dataloader): + for i, batch in enumerate(self.eval_dataloader): batch = to_device(batch, self.device) (input_ids, attention_mask, loss_mask, label, kl_input_ids, kl_attention_mask, kl_loss_mask) = ( batch["input_ids"],