feat(data, inference): 使用chatML格式

This commit is contained in:
2025-10-29 12:02:43 +08:00
parent 38b2725cd1
commit d94fc5a87a
8 changed files with 51 additions and 114 deletions
+2 -3
View File
@@ -68,11 +68,10 @@ class SftStrategy(BaseStrategy):
def compute_loss(self, batch: Dict[str, Tensor]) -> Tensor:
batch = move_to_device(batch, self.device)
input_ids, target_ids = batch["input_ids"], batch["target_ids"]
loss_mask, attn_mask = batch["loss_mask"], batch["attn_mask"]
input_ids, target_ids, loss_mask = batch["input_ids"], batch["target_ids"], batch["loss_mask"]
ignore_index = -100
logits = self.model(input_ids=input_ids, input_mask=attn_mask)["logits"]
logits = self.model(input_ids=input_ids)["logits"]
target_ids = target_ids.masked_fill(loss_mask == 0, ignore_index)
loss = F.cross_entropy(