#!/usr/bin/env bash
# Copy finished generation runs from the GPU machines into results/runs. Only the graded records and summaries come
# back (a few MB); the full answer texts (*.texts.jsonl.gz) stay on the machines.
#   bash scripts/pull_runs.sh trainer sampler
set -euo pipefail
cd "$(dirname "$0")/.."
mkdir -p results/runs
for h in "$@"; do
  rsync -azm --include '*/' --include 'merged.jsonl' --include 'summary.json' --include '*.done.json' \
    --exclude '*' "$h:/nvme/wsp-lab/runs/" results/runs/ </dev/null
  echo "pulled $h"
done