29 lines
1.1 KiB
Bash
Executable file
29 lines
1.1 KiB
Bash
Executable file
#!/bin/bash
|
|
# Wait for eval-suite-a, then run B and C sequentially.
|
|
# All output goes to docs/model-encyclopedia/benchmarks/_eval-*-stdout.log
|
|
set -u
|
|
MB=/var/home/lilith/Code/@applications/@model-boss
|
|
LOG_DIR=$MB/docs/model-encyclopedia/benchmarks
|
|
|
|
echo "[chain] $(date -Iseconds) waiting for eval-suite-a..."
|
|
while systemctl --user is-active eval-suite-a >/dev/null 2>&1; do
|
|
sleep 30
|
|
done
|
|
echo "[chain] $(date -Iseconds) eval-suite-a finished. Status: $(systemctl --user show -p Result --value eval-suite-a)"
|
|
|
|
# Suite B: code benchmarks
|
|
echo "[chain] $(date -Iseconds) launching suite B (code)..."
|
|
cd "$MB"
|
|
python3 -m tools.benchmark.cli code \
|
|
--models ministral-14b-reasoning,qwen3.6-27b,qwen3.6-35b-a3b,mistral-small-3.2-24b \
|
|
--samples 50 > "$LOG_DIR/_eval-suite-b-stdout.log" 2>&1
|
|
echo "[chain] $(date -Iseconds) suite B exit=$?"
|
|
|
|
# Suite C: vision describe
|
|
echo "[chain] $(date -Iseconds) launching suite C (describe)..."
|
|
python3 -m tools.benchmark.cli describe \
|
|
--models qwen3-vl-8b-instruct,gemma-4-31b \
|
|
--samples 15 > "$LOG_DIR/_eval-suite-c-stdout.log" 2>&1
|
|
echo "[chain] $(date -Iseconds) suite C exit=$?"
|
|
|
|
echo "[chain] $(date -Iseconds) all done."
|