code/slurm_logs/eval_after_sft.sbatch

#!/bin/bash
#SBATCH --job-name=vl
#SBATCH --partition=gpu-large
#SBATCH --qos=batch-long
#SBATCH --gres=gpu:1
#SBATCH --cpus-per-task=4
#SBATCH --mem=100G
#SBATCH --time=06:00:00
#SBATCH --output=/weka/s225250685/mats-tist/slurm_logs/eval_after_sft_%j.out

# Wait for all 5 SFT outputs to exist, then run 3 evaluation configs:
#   A. pass@1 greedy (T=0, K=1) — planner only
#   B. pass@8 multinomial (T=1.0, K=8) — planner only (oracle baseline, "no V+F")
#   C. pass@8 full pipeline (T=1.0, K=8) — planner + val-sel + val-cond + fixer
#   D. Selector EX on rollout C
#
# This produces the metrics the user wants:
#   - pass@1 (T=0): planner SFT quality
#   - pass@8 (T=1.0) no-VF: best-case oracle without help
#   - pass@8 (T=1.0) with VF: shows validator+fixer boost (key claim)
#   - Selector EX: final task accuracy

set -u
cd /weka/s225250685/mats-tist

set -a
source /weka/s225250685/mats-tist/.env
set +a

export HF_HOME=/weka/s225250685/Huggingface
export HF_HUB_CACHE=/weka/s225250685/Huggingface/hub
export DB_EXEC_API_DISABLE=1
export PYTHONNOUSERSITE=1
export NO_PROXY=localhost,127.0.0.1
export PYTHONPATH=/weka/s225250685/mats-tist
export TOKENIZERS_PARALLELISM=false

PY=/weka/s225250685/conda-envs/handbook/bin/python
VLLM=/weka/s225250685/conda-envs/handbook/bin/vllm
AH=/weka/s225250685/mats-tist/alignment-handbook

PLANNER=$AH/output/sft-planner-3B-griffith-v4
VAL_SEL=$AH/output/sft-validator-sel-llama1b-griffith-v5
VAL_COND=$AH/output/sft-validator-cond-llama1b-griffith-v5
FIXER=$AH/output/sft-fixer-llama1b-griffith-v5
SELECTOR=$AH/output/sft-selector-3B-griffith-v5

LOG=/weka/s225250685/mats-tist/slurm_logs/eval_after_sft_${SLURM_JOB_ID}.log
: > "$LOG"
nvidia-smi --query-gpu=name,memory.total --format=csv,noheader | tee -a "$LOG"

kill_vllm() {
    pkill -9 -f "vllm serve" 2>/dev/null || true
    pkill -9 -f "VLLM::EngineCore" 2>/dev/null || true
    sleep 5
}
trap kill_vllm EXIT

wait_url() {
    for i in {1..180}; do
        curl --noproxy '*' -fs "$1" >/dev/null 2>&1 && return 0; sleep 5
    done; echo "TIMEOUT: $1" | tee -a "$LOG"; return 1
}

##############################################
# WAIT for all 5 SFT outputs
##############################################
echo "==== [WAIT] checking SFT outputs ====" | tee -a "$LOG"
for i in $(seq 1 96); do
    ok=0
    for p in "$PLANNER" "$VAL_SEL" "$VAL_COND" "$FIXER" "$SELECTOR"; do
        [ -f "$p/config.json" ] && ok=$((ok+1))
    done
    echo "  [$i*5min] SFT outputs ready: $ok/5" | tee -a "$LOG"
    [ "$ok" -eq 5 ] && break
    sleep 300
done
for p in "$PLANNER" "$VAL_SEL" "$VAL_COND" "$FIXER" "$SELECTOR"; do
    [ -f "$p/config.json" ] || { echo "MISSING: $p" | tee -a "$LOG"; exit 1; }
done
echo "==== [WAIT] all SFT outputs found ====" | tee -a "$LOG"

##############################################
# A. pass@1 greedy (T=0, K=1) — planner only
##############################################
echo "==== [A] pass@1 greedy — planner only ====" | tee -a "$LOG"
kill_vllm
$VLLM serve "$PLANNER" --served-model-name planner --port 8100 \
    --dtype bfloat16 --gpu-memory-utilization 0.30 \
    --enforce-eager --max-model-len 8192 > "${LOG}.p" 2>&1 &
wait_url http://localhost:8100/v1/models && echo "  planner READY" | tee -a "$LOG"

OUT_A=eval_results/sft_phase1_passAt1_greedy_bird_dev.jsonl
rm -f "$OUT_A"
$PY scripts/run_pipeline_rollouts.py \
    --input_file data/sft_bird_with_evidence_dev_text2sql.json \
    --output_file "$OUT_A" \
    --planner_host http://localhost:8100 --planner_format qwen \
    --validator_host none --fixer_host none \
    --griffith_prompts \
    --K 1 --K_val 1 --K_fix 1 \
    --temperature 0.0 --top_p 1.0 \
    --max_planner_tokens 1024 \
    --max_questions -1 --n_threads 4 2>&1 | tee -a "$LOG"
$PY scripts/compute_bestofn_metrics.py "$OUT_A" sft_passAt1_greedy 2>&1 | tee -a "$LOG"

##############################################
# B. pass@8 multinomial (T=1.0, K=8) — planner only (oracle, no V+F)
##############################################
echo "==== [B] pass@8 multinomial — planner only ====" | tee -a "$LOG"
# Reuse same planner endpoint
OUT_B=eval_results/sft_phase1_passAt8_noVF_bird_dev.jsonl
rm -f "$OUT_B"
$PY scripts/run_pipeline_rollouts.py \
    --input_file data/sft_bird_with_evidence_dev_text2sql.json \
    --output_file "$OUT_B" \
    --planner_host http://localhost:8100 --planner_format qwen \
    --validator_host none --fixer_host none \
    --griffith_prompts \
    --K 8 --K_val 1 --K_fix 1 \
    --temperature 1.0 --top_p 0.9 \
    --max_planner_tokens 1024 \
    --max_questions -1 --n_threads 4 2>&1 | tee -a "$LOG"
$PY scripts/compute_bestofn_metrics.py "$OUT_B" sft_passAt8_noVF 2>&1 | tee -a "$LOG"

##############################################
# C. pass@8 full pipeline (T=1.0, K=8) — planner + V + F
##############################################
echo "==== [C] pass@8 full pipeline — planner + V + F ====" | tee -a "$LOG"
$VLLM serve "$VAL_SEL" --served-model-name validator_sel --port 8101 \
    --dtype bfloat16 --gpu-memory-utilization 0.10 \
    --enforce-eager --max-model-len 6144 > "${LOG}.vs" 2>&1 &
wait_url http://localhost:8101/v1/models && echo "  val-sel READY" | tee -a "$LOG"

$VLLM serve "$VAL_COND" --served-model-name validator_cond --port 8104 \
    --dtype bfloat16 --gpu-memory-utilization 0.10 \
    --enforce-eager --max-model-len 6144 > "${LOG}.vc" 2>&1 &
wait_url http://localhost:8104/v1/models && echo "  val-cond READY" | tee -a "$LOG"

$VLLM serve "$FIXER" --served-model-name fixer --port 8102 \
    --dtype bfloat16 --gpu-memory-utilization 0.12 \
    --enforce-eager --max-model-len 6144 > "${LOG}.f" 2>&1 &
wait_url http://localhost:8102/v1/models && echo "  fixer READY" | tee -a "$LOG"

OUT_C=eval_results/sft_phase1_passAt8_withVF_bird_dev.jsonl
rm -f "$OUT_C"
$PY scripts/run_pipeline_rollouts.py \
    --input_file data/sft_bird_with_evidence_dev_text2sql.json \
    --output_file "$OUT_C" \
    --planner_host http://localhost:8100 --planner_format qwen \
    --validator_host none \
    --validator_sel_host http://localhost:8101 \
    --validator_cond_host http://localhost:8104 \
    --fixer_host http://localhost:8102 \
    --fixer_gate_exec_ok \
    --griffith_prompts \
    --K 8 --K_val 1 --K_fix 1 \
    --temperature 1.0 --top_p 0.9 \
    --max_planner_tokens 1024 --max_validator_tokens 384 --max_fixer_tokens 512 \
    --max_questions -1 --n_threads 4 2>&1 | tee -a "$LOG"
$PY scripts/compute_bestofn_metrics.py "$OUT_C" sft_passAt8_withVF 2>&1 | tee -a "$LOG"

##############################################
# D. Selector EX on rollout C
##############################################
echo "==== [D] Selector EX on rollout C ====" | tee -a "$LOG"
kill_vllm
$VLLM serve "$SELECTOR" --served-model-name selector --port 8103 \
    --dtype bfloat16 --gpu-memory-utilization 0.85 \
    --enforce-eager --max-model-len 4096 > "${LOG}.sel" 2>&1 &
wait_url http://localhost:8103/v1/models && echo "  selector READY" | tee -a "$LOG"

$PY scripts/compute_bestofn_with_selector.py \
    "$OUT_C" sft_selectorEX \
    --selector_host http://localhost:8103 --row_preview 2>&1 | tee -a "$LOG"

##############################################
# SUMMARY
##############################################
echo "" | tee -a "$LOG"
echo "===========================================" | tee -a "$LOG"
echo "PHASE 1 SFT — EVALUATION SUMMARY" | tee -a "$LOG"
echo "===========================================" | tee -a "$LOG"
grep -E "===.*passAt|===.*selectorEX|pass@|greedy|oracle|selector" "$LOG" | tail -30
echo "==== ALL_DONE ====" | tee -a "$LOG"