CI: stop and retry dataset preparation when a download stalls

.github/workflows/pr-evaluation.yml

@@ -258,7 +258,7 @@ jobs:
           # Uses base repo's script (safe - not from PR)
           echo "Preparing dataset..."
           mkdir -p dataset
-          uv run python scripts/process_datasets/prep_datasets.py
+          scripts/process_datasets/prepare_datasets.sh
 
       - name: Enable Swap Space
         run: |

.github/workflows/regrade-all.yml

@@ -88,7 +88,7 @@ jobs:
         run: |
           set -euo pipefail
           mkdir -p dataset
-          uv run python scripts/process_datasets/prep_datasets.py
+          scripts/process_datasets/prepare_datasets.sh
 
       - name: Enable swap space
         run: |

.github/workflows/sync-website.yml

@@ -41,7 +41,7 @@ jobs:
       - name: Prepare dataset
         env:
           ROUTERARENA_DATASET_DIR: ${{ github.workspace }}/dataset
-        run: uv run python scripts/process_datasets/prep_datasets.py
+        run: scripts/process_datasets/prepare_datasets.sh
 
       - name: Checkout website repo
         uses: actions/checkout@v5

scripts/process_datasets/prepare_datasets.sh

@@ -0,0 +1,26 @@
+#!/usr/bin/env bash
+# SPDX-FileCopyrightText: Copyright contributors to the RouterArena project
+# SPDX-License-Identifier: Apache-2.0
+#
+# Run prep_datasets.py with a time limit, retrying a few times. The Hugging Face
+# downloads it makes occasionally stall without failing, which used to hang CI jobs
+# until their own timeout; a stalled attempt is now stopped and retried.
+#
+# Usage: scripts/process_datasets/prepare_datasets.sh
+#   PREP_TIMEOUT  seconds per attempt (default 900)
+#   PREP_ATTEMPTS number of attempts (default 3)
+set -euo pipefail
+
+timeout_s="${PREP_TIMEOUT:-900}"
+attempts="${PREP_ATTEMPTS:-3}"
+for attempt in $(seq 1 "$attempts"); do
+  if timeout --kill-after=30 "$timeout_s" uv run python scripts/process_datasets/prep_datasets.py; then
+    exit 0
+  fi
+  echo "Dataset preparation attempt $attempt of $attempts failed or timed out." >&2
+  if [[ "$attempt" -lt "$attempts" ]]; then
+    sleep 15
+  fi
+done
+echo "Dataset preparation failed after $attempts attempts." >&2
+exit 1