#!/bin/bash
# Evaluate a model on one benchmark. No API key needed except for the LLM-graded tasks
# (arenahardwriting, healthbench), which need OPENAI_API_KEY in .env.
#
#   sbatch scripts/ptb_eval.sbatch                                  # full gsm8k baseline
#   LIMIT=20 sbatch scripts/ptb_eval.sbatch                         # quick sanity check
#   MODEL=<run dir>/final_model OUT=<...> sbatch scripts/ptb_eval.sbatch
#SBATCH --job-name=ptb-eval
#SBATCH --partition=h100
#SBATCH --account=midpri
#SBATCH --qos=midpri
#SBATCH --gres=gpu:1
#SBATCH --time=04:00:00
#SBATCH --cpus-per-task=16
#SBATCH --mem=256G
#SBATCH --output=logs/ptb-eval-%j.out

set -euo pipefail
unset http_proxy https_proxy HTTP_PROXY HTTPS_PROXY no_proxy NO_PROXY X2P_PROXY_URL

TASK="${TASK:-gsm8k}"
MODEL="${MODEL:-Qwen/Qwen3-1.7B-Base}"
LIMIT="${LIMIT:--1}"            # -1 = the full benchmark
OUT="${OUT:-/home/jiaruiliu/ptb-data/baselines/${TASK}_qwen3-1.7b}"

cd /home/jiaruiliu/local/git-repos/auto-impl
export PYTHONPATH=src

echo "node=$(hostname)"; nvidia-smi -L
python3 -m posttrainbench eval --task "$TASK" --model "$MODEL" --limit "$LIMIT" --out-dir "$OUT"