From 6cb5902eb6df5e2875ea3d73b6d4708ba58ede3d Mon Sep 17 00:00:00 2001
From: Josh Creek <8179928+jcreek@users.noreply.github.com>
Date: Mon, 20 Jul 2026 07:18:13 +0100
Subject: [PATCH] feat(*): Add long training run script

---
 training/next_long_run.sh | 17 +++++++++++++++++
 1 file changed, 17 insertions(+)
 create mode 100755 training/next_long_run.sh

diff --git a/training/next_long_run.sh b/training/next_long_run.sh
new file mode 100755
index 00000000..48c1c1e9
--- /dev/null
+++ b/training/next_long_run.sh
@@ -0,0 +1,17 @@
+#!/usr/bin/env bash
+# The overnight version of next_run.sh: same zero-thought kickoff (auto-names
+# the run, resumes from the newest committed final.zip, same standing flags),
+# but sized for a ~12-hour run instead of ~2.5 hours.
+#
+#   --timesteps 90000000: runs 05-07 all trained at ~7.5M steps/hour on the
+#     box (measured from the tfevents wall times), so 90M new steps ~= 12h.
+#     On resume SB3 treats the budget as additional steps, same as the 20M runs.
+#   --checkpoint-every 500000: at the default 100k cadence a 90M run would
+#     commit ~900 checkpoint zips (~190MB); 500k keeps it to ~180 (~38MB),
+#     about the same repo growth as a normal run.
+#
+# Usage: ./next_long_run.sh [extra train.py args...]
+# Extras still win on conflict, e.g. ./next_long_run.sh --timesteps 60000000
+# for a shorter overnight run.
+set -euo pipefail
+exec "$(dirname "$0")/next_run.sh" --timesteps 90000000 --checkpoint-every 500000 "$@"