diff --git a/soup_cli/commands/generate.py b/soup_cli/commands/generate.py
index a057553..59e201a 100644
--- a/soup_cli/commands/generate.py
+++ b/soup_cli/commands/generate.py
@@ -454,7 +454,10 @@ def _generate_server(
         )
 
     data = response.json()
-    content = data["choices"][0]["message"]["content"]
+    try:
+        content = data["choices"][0]["message"]["content"]
+    except (KeyError, IndexError, TypeError) as exc:
+        raise ValueError(f"Unexpected server response format: {exc}") from exc
 
     return _parse_json_array(content)
 
diff --git a/soup_cli/data/loader.py b/soup_cli/data/loader.py
index 16f7235..7ad8c98 100644
--- a/soup_cli/data/loader.py
+++ b/soup_cli/data/loader.py
@@ -198,8 +198,7 @@ def _validate_audio_files(data: list[dict], audio_dir: Path) -> list[dict]:
         if not str(resolved).startswith(str(resolved_base)):
             traversal += 1
             continue
-        row["audio"] = str(resolved)
-        valid.append(row)
+        valid.append({**row, "audio": str(resolved)})
 
     if missing > 0:
         console.print(f"[yellow]Warning: {missing} rows skipped (missing audio path)[/]")
diff --git a/soup_cli/trainer/sft.py b/soup_cli/trainer/sft.py
index d283d04..a83bfa0 100644
--- a/soup_cli/trainer/sft.py
+++ b/soup_cli/trainer/sft.py
@@ -516,6 +516,12 @@ class SFTTrainerWrapper:
                 result["sampling_rate"] = sampling_rate
             return result
 
+        if not dataset["train"]:
+            raise ValueError(
+                "Audio training dataset is empty after validation. "
+                "Check audio file paths and audio_dir."
+            )
+
         remove_cols = ["messages", "audio"]
         train_ds = Dataset.from_list(dataset["train"]).map(
             load_and_format_audio,
diff --git a/soup_cli/utils/quality.py b/soup_cli/utils/quality.py
index aafca17..f72d9a8 100644
--- a/soup_cli/utils/quality.py
+++ b/soup_cli/utils/quality.py
@@ -78,6 +78,11 @@ def compute_perplexity_scores(
                 ppl = math.exp(min(loss_val.item(), 100))  # cap to avoid overflow
                 scores.append(ppl)
 
+    # Cleanup GPU memory
+    del model
+    if torch.cuda.is_available():
+        torch.cuda.empty_cache()
+
     return scores