diff --git a/Dockerfile b/Dockerfile
index c13ecb9ef..e59376826 100644
--- a/Dockerfile
+++ b/Dockerfile
@@ -2,8 +2,6 @@ FROM python:3.11-slim-bullseye
 
 # Allow statements and log messages to immediately appear in the Knative logs
 ENV PYTHONUNBUFFERED True
-# Fixes "invalid instruction" runtime error on AMD machines (specifically HF Upgraded CPU Space).
-ENV HNSWLIB_NO_NATIVE 1
 
 # Adds GCC and other build tools so we can compile hnswlib and other native/C++ deps.
 RUN apt-get update --fix-missing && apt-get install -y --fix-missing build-essential && \
diff --git a/lilac/data/clustering.py b/lilac/data/clustering.py
index e88f96c0c..cb8ca09f3 100644
--- a/lilac/data/clustering.py
+++ b/lilac/data/clustering.py
@@ -449,29 +449,43 @@ def compute_category_titles(items: Iterator[Item]) -> Iterator[Item]:
       output_path=cluster_output_path,
       sort_by=(*cluster_output_path, CATEGORY_ID),
       overwrite=True,
-      schema=field(
-        fields={
-          CLUSTER_ID: field('int32', categorical=True),
-          CLUSTER_MEMBERSHIP_PROB: 'float32',
-          CLUSTER_TITLE: 'string',
-          CATEGORY_ID: field('int32', categorical=True),
-          CATEGORY_MEMBERSHIP_PROB: 'float32',
-          CATEGORY_TITLE: 'string',
-        },
-        cluster=ClusterInfo(
-          min_cluster_size=min_cluster_size,
-          use_garden=use_garden,
-          input_path=(get_callable_name(input_fn_or_path),) if callable(input_fn_or_path) else path,
-          input_format_selector=ClusterInputFormatSelectorInfo(
-            format=manifest.dataset_format.name,
-            selector=dataset_format_input_selector.name,
-          )
-          if dataset_format_input_selector and manifest.dataset_format
-          else None,
-        ),
-      ),
     )
 
+  def drop_temp_text_column(items: Iterator[Item]) -> Iterator[Item]:
+    for item in items:
+      if TEXT_COLUMN in item:
+        del item[TEXT_COLUMN]
+      yield item
+
+  # Drop the temporary newline-concatenated text column and write the final output.
+  dataset.transform(
+    drop_temp_text_column,
+    input_path=cluster_output_path,
+    output_path=cluster_output_path,
+    overwrite=True,
+    schema=field(
+      fields={
+        CLUSTER_ID: field('int32', categorical=True),
+        CLUSTER_MEMBERSHIP_PROB: 'float32',
+        CLUSTER_TITLE: 'string',
+        CATEGORY_ID: field('int32', categorical=True),
+        CATEGORY_MEMBERSHIP_PROB: 'float32',
+        CATEGORY_TITLE: 'string',
+      },
+      cluster=ClusterInfo(
+        min_cluster_size=min_cluster_size,
+        use_garden=use_garden,
+        input_path=(get_callable_name(input_fn_or_path),) if callable(input_fn_or_path) else path,
+        input_format_selector=ClusterInputFormatSelectorInfo(
+          format=manifest.dataset_format.name,
+          selector=dataset_format_input_selector.name,
+        )
+        if dataset_format_input_selector and manifest.dataset_format
+        else None,
+      ),
+    ),
+  )
+
 
 def _hdbscan_cluster(
   docs: Iterator[str],
diff --git a/lilac/hf_docker/Dockerfile b/lilac/hf_docker/Dockerfile
index 4dddfa927..e9b697714 100644
--- a/lilac/hf_docker/Dockerfile
+++ b/lilac/hf_docker/Dockerfile
@@ -2,6 +2,8 @@ FROM python:3.11-slim-bullseye
 
 # Allow statements and log messages to immediately appear in the Knative logs
 ENV PYTHONUNBUFFERED True
+# Fixes "invalid instruction" runtime error on AMD machines (specifically HF Upgraded CPU Space).
+ENV HNSWLIB_NO_NATIVE 1
 
 # Adds GCC and other build tools so we can compile hnswlib and other native/C++ deps.
 RUN apt-get update --fix-missing && apt-get install -y --fix-missing build-essential && \
diff --git a/lilac/router_dataset_signals.py b/lilac/router_dataset_signals.py
index d959afa30..7c90335cc 100644
--- a/lilac/router_dataset_signals.py
+++ b/lilac/router_dataset_signals.py
@@ -1,5 +1,4 @@
 """Routing endpoints for running signals on datasets."""
-from threading import Thread
 from typing import Annotated, Optional
 
 from fastapi import APIRouter, HTTPException
@@ -122,8 +121,7 @@ def run() -> None:
       task_id=task_id,
     )
 
-  thread = Thread(target=run, daemon=True)
-  thread.start()
+  launch_task(task_id, run)
   return ClusterResponse(task_id=task_id)