gpudrive-main/scripts/evaluate_llm_cases_deepseek.sbatch

#!/bin/bash
#SBATCH --partition=gre
#SBATCH --nodes=1
#SBATCH --ntasks=1
#SBATCH --time=24:00:00
#SBATCH --job-name=gpudrive_eval_agent
#SBATCH --output=%x-%A_%a.out
#SBATCH --mem=64G
#SBATCH --cpus-per-task=8

set -euo pipefail

# Single shard:
#   sbatch scripts/evaluate_llm_cases_deepseek.sbatch
#
# Multiple shards:
#   sbatch --array=0-3 scripts/evaluate_llm_cases_deepseek.sbatch

PROJECT_ROOT="${PROJECT_ROOT:-/mingli01/project/zihanliu/gpudrive-main}"
CONDA_ENV_NAME="${CONDA_ENV_NAME:-gpudrive}"

CASES_BASE="${CASES_BASE:-search_outputs/llm_cases}"
OUTPUT_BASE="${OUTPUT_BASE:-search_outputs/evaluator_agent}"
RULES_PATH="${RULES_PATH:-rules/longtail_llm_case_evaluation.md}"

SHARD_ID="${SLURM_ARRAY_TASK_ID:-${SHARD_ID:-0}}"
SHARD_PADDED="$(printf "%03d" "${SHARD_ID}")"

CASES="${CASES:-${CASES_BASE}/shard_${SHARD_PADDED}/llm_cases.jsonl}"
OUTPUT_DIR="${OUTPUT_DIR:-${OUTPUT_BASE}/shard_${SHARD_PADDED}}"

START_INDEX="${START_INDEX:-0}"
MAX_CASES="${MAX_CASES:-0}"
BATCH_SIZE="${BATCH_SIZE:-1}"
ACCEPT_THRESHOLD="${ACCEPT_THRESHOLD:-70.0}"
HIGH_VALUE_THRESHOLD="${HIGH_VALUE_THRESHOLD:-82.0}"
DRY_RUN="${DRY_RUN:-0}"
RESUME="${RESUME:-1}"
WRITE_PARTIAL_SUMMARY="${WRITE_PARTIAL_SUMMARY:-1}"
SAVE_PROMPTS="${SAVE_PROMPTS:-0}"
SAVE_RAW_RESPONSES="${SAVE_RAW_RESPONSES:-1}"
FALLBACK_ON_PARSE_ERROR="${FALLBACK_ON_PARSE_ERROR:-1}"
RETRY_SINGLE_ON_PARSE_ERROR="${RETRY_SINGLE_ON_PARSE_ERROR:-1}"

DEEPSEEK_BASE_URL="${DEEPSEEK_BASE_URL:-https://api.deepseek.com}"
DEEPSEEK_MODEL="${DEEPSEEK_MODEL:-deepseek-v4-flash}"
TEMPERATURE="${TEMPERATURE:-0.0}"
MAX_TOKENS="${MAX_TOKENS:-8192}"
TIMEOUT_SEC="${TIMEOUT_SEC:-120}"
RETRIES="${RETRIES:-3}"
SLEEP_SEC="${SLEEP_SEC:-0.5}"

CURRENT_COLLISION_WEIGHT="${CURRENT_COLLISION_WEIGHT:-0.0}"
CURRENT_GOAL_WEIGHT="${CURRENT_GOAL_WEIGHT:-2.0}"
CURRENT_OFFROAD_WEIGHT="${CURRENT_OFFROAD_WEIGHT:--0.4}"
CURRENT_RISK_AGENTS_PER_WORLD="${CURRENT_RISK_AGENTS_PER_WORLD:-3}"
CURRENT_NORMAL_MODE="${CURRENT_NORMAL_MODE:-policy}"
CURRENT_DETERMINISTIC="${CURRENT_DETERMINISTIC:-0}"
CURRENT_NORMAL_STYLE="${CURRENT_NORMAL_STYLE:-balanced}"
CURRENT_RISK_STYLE="${CURRENT_RISK_STYLE:-risk_taker}"

module load conda/3
source "$(conda info --base)/etc/profile.d/conda.sh"
conda activate "${CONDA_ENV_NAME}"

export PYTHONNOUSERSITE=1
export PYTHONUNBUFFERED=1
export OMP_NUM_THREADS="${OMP_NUM_THREADS:-1}"
export MKL_NUM_THREADS="${MKL_NUM_THREADS:-1}"
export DEEPSEEK_BASE_URL
export DEEPSEEK_MODEL

cd "${PROJECT_ROOT}"
export PYTHONPATH="${PROJECT_ROOT}:${PYTHONPATH:-}"

if [[ ! -f "${CASES}" ]]; then
  echo "ERROR: CASES not found: ${CASES}"
  exit 1
fi
if [[ ! -f "${RULES_PATH}" ]]; then
  echo "ERROR: RULES_PATH not found: ${RULES_PATH}"
  exit 1
fi
if [[ "${DRY_RUN}" != "1" && -z "${DEEPSEEK_API_KEY:-}" ]]; then
  echo "ERROR: DEEPSEEK_API_KEY is not set. Use DRY_RUN=1 for rule-only evaluation."
  exit 1
fi

mkdir -p "${OUTPUT_DIR}"

echo "SLURM_JOB_ID=${SLURM_JOB_ID:-unknown}"
echo "SLURM_ARRAY_TASK_ID=${SLURM_ARRAY_TASK_ID:-none}"
echo "PROJECT_ROOT=${PROJECT_ROOT}"
echo "SHARD_ID=${SHARD_ID}"
echo "CASES=${CASES}"
echo "RULES_PATH=${RULES_PATH}"
echo "OUTPUT_DIR=${OUTPUT_DIR}"
echo "START_INDEX=${START_INDEX}"
echo "MAX_CASES=${MAX_CASES}"
echo "BATCH_SIZE=${BATCH_SIZE}"
echo "DRY_RUN=${DRY_RUN}"
echo "RESUME=${RESUME}"
echo "WRITE_PARTIAL_SUMMARY=${WRITE_PARTIAL_SUMMARY}"
echo "SAVE_RAW_RESPONSES=${SAVE_RAW_RESPONSES}"
echo "FALLBACK_ON_PARSE_ERROR=${FALLBACK_ON_PARSE_ERROR}"
echo "RETRY_SINGLE_ON_PARSE_ERROR=${RETRY_SINGLE_ON_PARSE_ERROR}"
echo "DEEPSEEK_MODEL=${DEEPSEEK_MODEL}"

python -u search/evaluate_llm_cases_deepseek.py \
  --cases "${CASES}" \
  --rules "${RULES_PATH}" \
  --output-dir "${OUTPUT_DIR}" \
  --start-index "${START_INDEX}" \
  --max-cases "${MAX_CASES}" \
  --batch-size "${BATCH_SIZE}" \
  --accept-threshold "${ACCEPT_THRESHOLD}" \
  --high-value-threshold "${HIGH_VALUE_THRESHOLD}" \
  --dry-run "${DRY_RUN}" \
  --resume "${RESUME}" \
  --write-partial-summary "${WRITE_PARTIAL_SUMMARY}" \
  --save-prompts "${SAVE_PROMPTS}" \
  --save-raw-responses "${SAVE_RAW_RESPONSES}" \
  --fallback-on-parse-error "${FALLBACK_ON_PARSE_ERROR}" \
  --retry-single-on-parse-error "${RETRY_SINGLE_ON_PARSE_ERROR}" \
  --base-url "${DEEPSEEK_BASE_URL}" \
  --model "${DEEPSEEK_MODEL}" \
  --temperature "${TEMPERATURE}" \
  --max-tokens "${MAX_TOKENS}" \
  --timeout-sec "${TIMEOUT_SEC}" \
  --retries "${RETRIES}" \
  --sleep-sec "${SLEEP_SEC}" \
  --current-collision-weight "${CURRENT_COLLISION_WEIGHT}" \
  --current-goal-weight "${CURRENT_GOAL_WEIGHT}" \
  --current-offroad-weight "${CURRENT_OFFROAD_WEIGHT}" \
  --current-risk-agents-per-world "${CURRENT_RISK_AGENTS_PER_WORLD}" \
  --current-normal-mode "${CURRENT_NORMAL_MODE}" \
  --current-deterministic "${CURRENT_DETERMINISTIC}" \
  --current-normal-style "${CURRENT_NORMAL_STYLE}" \
  --current-risk-style "${CURRENT_RISK_STYLE}"

echo "Evaluator Agent finished."
echo "Outputs:"
find "${OUTPUT_DIR}" -maxdepth 1 -type f | sort