diff --git a/model2vec/train/classifier.py b/model2vec/train/classifier.py
index 93c3cd5..1f98000 100644
--- a/model2vec/train/classifier.py
+++ b/model2vec/train/classifier.py
@@ -22,9 +22,9 @@
 
 
 def _classifier_metrics(head_out: torch.Tensor, y: torch.Tensor, loss: torch.Tensor) -> dict[str, float]:
-    """Validation metrics for single-label classification: loss and accuracy."""
+    """Metrics for single-label classification: loss and accuracy."""
     accuracy = (head_out.argmax(dim=1) == y).float().mean()
-    return {"val_loss": loss.item(), "val_accuracy": accuracy.item()}
+    return {"loss": loss.item(), "accuracy": accuracy.item()}
 
 
 def _compute_accuracy(y_true: torch.Tensor, y_pred: torch.Tensor) -> float:
@@ -36,10 +36,10 @@ def _compute_accuracy(y_true: torch.Tensor, y_pred: torch.Tensor) -> float:
 
 
 def _multilabel_classifier_metrics(head_out: torch.Tensor, y: torch.Tensor, loss: torch.Tensor) -> dict[str, float]:
-    """Validation metrics for multi-label classification: loss and Jaccard accuracy."""
+    """Metrics for multi-label classification: loss and Jaccard accuracy."""
     preds = (torch.sigmoid(head_out) > 0.5).float()
     accuracy = _compute_accuracy(y, preds)
-    return {"val_loss": loss.item(), "val_accuracy": accuracy}
+    return {"loss": loss.item(), "accuracy": accuracy}
 
 
 class StaticModelForClassification(BaseFinetuneable):
diff --git a/model2vec/train/trainer.py b/model2vec/train/trainer.py
index 2c45261..6fe0493 100644
--- a/model2vec/train/trainer.py
+++ b/model2vec/train/trainer.py
@@ -1,7 +1,7 @@
 from __future__ import annotations
 
 import copy
-from collections import defaultdict
+from collections import defaultdict, deque
 from collections.abc import Callable
 
 import torch
@@ -12,11 +12,12 @@
 MetricsFn = Callable[[torch.Tensor, torch.Tensor, torch.Tensor], dict[str, float]]
 
 _UNBOUNDED_MAX_EPOCHS = 9999
+_TRAIN_METRICS_WINDOW = 50
 
 
 def default_metrics(head_out: torch.Tensor, y: torch.Tensor, loss: torch.Tensor) -> dict[str, float]:
-    """Validation metrics for tasks that only track loss (used for early stopping on val_loss)."""
-    return {"val_loss": loss.item()}
+    """Metrics for tasks that only track loss (used for early stopping on val_loss)."""
+    return {"loss": loss.item()}
 
 
 class EarlyStopper:
@@ -82,7 +83,7 @@ def _run_validation(
         head_out = model(x)
         loss = loss_function(head_out, y)
         for key, value in compute_metrics(head_out, y, loss).items():
-            weighted_sums[key] += value * batch_size
+            weighted_sums[f"val_{key}"] += value * batch_size
         total_samples += batch_size
     model.train()
     return {key: total / total_samples for key, total in weighted_sums.items()}
@@ -109,7 +110,7 @@ def run_training_loop(  # noqa: C901
     :param model: The model to train, called as `head_out = model(x)`.
     :param loss_function: Computes the training and validation loss from `(head_out, y)`.
     :param learning_rate: The Adam learning rate.
-    :param val_metric: The metric key (returned by `compute_metrics`) used for early stopping.
+    :param val_metric: The metric key used for early stopping: a key returned by `compute_metrics`, prefixed with `val_`.
     :param early_stopping_direction: Either "min" or "max", the direction of improvement for `val_metric`.
     :param train_loader: The training data loader.
     :param val_loader: The validation data loader.
@@ -120,7 +121,7 @@ def run_training_loop(  # noqa: C901
     :param device: The device to train on.
     :param val_check_interval: If set, validate every this many training steps.
     :param check_val_every_epoch: If set, validate every this many epochs.
-    :param compute_metrics: Computes validation metrics from `(head_out, y, loss)`. Defaults to just `val_loss`.
+    :param compute_metrics: Computes unprefixed metrics from `(head_out, y, loss)`. Defaults to just `loss`.
     :return: The model's state dict from the validation check with the best `val_metric`.
     """
     model.to(device)
@@ -149,6 +150,7 @@ def run_training_loop(  # noqa: C901
     current_epoch = 0
     global_step = 0
     postfix: dict[str, str] = {}
+    train_metric_windows: dict[str, deque[float]] = defaultdict(lambda: deque(maxlen=_TRAIN_METRICS_WINDOW))
     latest_val_loss: float | None = None
 
     def validate_and_checkpoint() -> bool:
@@ -176,7 +178,12 @@ def validate_and_checkpoint() -> bool:
                 optimizer.step()
                 global_step += 1
 
-                postfix["train_loss"] = f"{loss.item():.4f}"
+                with torch.no_grad():
+                    train_metrics = compute_metrics(head_out, y, loss)
+                for key, value in train_metrics.items():
+                    train_metric_windows[f"train_{key}"].append(value)
+                for key, window in train_metric_windows.items():
+                    postfix[key] = f"{sum(window) / len(window):.4f}"
                 pbar.set_postfix(postfix)
 
                 if val_check_interval is not None and global_step % val_check_interval == 0:
diff --git a/tests/test_trainable.py b/tests/test_trainable.py
index 69f6be2..0bdafa4 100644
--- a/tests/test_trainable.py
+++ b/tests/test_trainable.py
@@ -936,7 +936,7 @@ def test_run_training_loop_mid_epoch_early_stop() -> None:
         device=resolve_device("cpu"),
         val_check_interval=1,
         check_val_every_epoch=None,
-        compute_metrics=lambda head_out, y, loss: {"val_loss": 1.0},
+        compute_metrics=lambda head_out, y, loss: {"loss": 1.0},
     )
     assert set(state_dict) == set(model.state_dict())