code/slurm_logs/mega_orpo_v3.sbatch

#!/bin/bash
#SBATCH --job-name=vl
#SBATCH --partition=gpu-large
#SBATCH --qos=batch-long
#SBATCH --gres=gpu:1
#SBATCH --cpus-per-task=2
#SBATCH --mem=80G
#SBATCH --time=12:00:00
#SBATCH --output=/weka/s225250685/mats-tist/slurm_logs/orpo_v3_%j.out

# ============================================================
# ORPO v3 — fixes over v2:
#   1. Separate validator-sel and validator-cond ORPO models
#      (cleaner training signal vs combined model)
#   2. Prompt format aligned with inference exactly
#      (field labels: "database schema:", "External knowledge:",
#       "Generated SQL query:", "Execution response:")
#   3. Semantic fixer preserve pairs use SAME prompt as inference
#      (not PRESERVE_PROMPT which never appears at inference)
#   4. Preserve rejected = cross-question wrong SQL (valid ORPO contrast)
#   5. Exec-error fixer uses EXEC_FIXER_PROMPT at inference
#      (no validator critique, "Failed SQL" / "Execution error" labels)
#   6. 3 epochs for validators (vs 2 in v2)
# ============================================================

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

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
ACCEL=/weka/s225250685/conda-envs/handbook/bin/accelerate
AH=/weka/s225250685/mats-tist/alignment-handbook

PLANNER=$AH/output/planner-iter2-collab-3B
SEL_V2=$AH/output/selector-3B-v2-rows
VAL_SEL=$AH/output/validator-sel-v4-orpo
VAL_COND=$AH/output/validator-cond-v4-orpo
FIXER_V2=$AH/output/fixer-v2-1.5B-execerr-orpo-expanded
SEM_FIXER=$AH/output/semantic-fixer-v3-1.5B-orpo-nogate

LOG=/weka/s225250685/mats-tist/slurm_logs/orpo_v3_${SLURM_JOB_ID}.log
: > "$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
    return 1
}

##############################################
# STAGE A: Build ORPO training data
##############################################
echo "==== [A] building ORPO training data ====" | tee -a "$LOG"

echo "  [A1] exec-error fixer data..." | tee -a "$LOG"
$PY scripts/build_fixer_v2_execerr.py 2>&1 | tee -a "$LOG"
[ -L data/hf_fixer_v2_execerr_expanded ] || ln -sf hf_fixer_v2_execerr data/hf_fixer_v2_execerr_expanded

echo "  [A2] validator v4 ORPO data (aligned prompts)..." | tee -a "$LOG"
$PY scripts/build_validator_v4_orpo.py 2>&1 | tee -a "$LOG"

# Split combined validator data into separate sel and cond datasets
echo "  [A2b] splitting validator data into sel/cond..." | tee -a "$LOG"
$PY - << 'PYEOF'
from datasets import load_from_disk, DatasetDict, Dataset
d = load_from_disk("data/hf_validator_v4_orpo")
for role in ["sel", "cond"]:
    tag = "select" if role == "sel" else "condition"
    for split in ["train_dpo", "test_dpo"]:
        ds = d[split].filter(lambda x: x["role"] == tag)
        print(f"  {role} {split}: {len(ds)}")
    DatasetDict({
        "train_dpo": d["train_dpo"].filter(lambda x: x["role"] == tag),
        "test_dpo":  d["test_dpo"].filter(lambda x: x["role"] == tag),
    }).save_to_disk(f"data/hf_validator_v4_{role}_orpo")
    print(f"  saved → data/hf_validator_v4_{role}_orpo")
PYEOF

echo "  [A3] semantic fixer v3 data (fixed preserve pairs)..." | tee -a "$LOG"
$PY scripts/build_semantic_fixer_v3.py 2>&1 | tee -a "$LOG"

echo "==== [A] data build done ====" | tee -a "$LOG"

##############################################
# STAGE B: ORPO validator-sel (0.5B) — skip if already trained
##############################################
if [ -f "$VAL_SEL/config.json" ]; then
    echo "==== [B] validator-sel already exists, skipping ====" | tee -a "$LOG"
else
    echo "==== [B] ORPO validator-sel (0.5B) ====" | tee -a "$LOG"
    cd $AH
    PYTHONPATH=src/ ACCELERATE_LOG_LEVEL=info $ACCEL launch \
        --main_process_port 29621 \
        --config_file recipes/accelerate_configs/single_gpu0_local.yaml \
        scripts/run_orpo.py \
        recipes/scaleup-3stage/orpo-validator-sel-v4.yaml 2>&1 | tee -a "$LOG"
    cd /weka/s225250685/mats-tist
    if [ ! -f "$VAL_SEL/config.json" ]; then
        echo "VALIDATOR-SEL ORPO FAILED" | tee -a "$LOG"; exit 1
    fi
fi
echo "==== [B] validator-sel ready: $VAL_SEL ====" | tee -a "$LOG"

##############################################
# STAGE C: ORPO validator-cond (0.5B) — skip if already trained
##############################################
if [ -f "$VAL_COND/config.json" ]; then
    echo "==== [C] validator-cond already exists, skipping ====" | tee -a "$LOG"
else
    echo "==== [C] ORPO validator-cond (0.5B) ====" | tee -a "$LOG"
    cd $AH
    PYTHONPATH=src/ ACCELERATE_LOG_LEVEL=info $ACCEL launch \
        --main_process_port 29622 \
        --config_file recipes/accelerate_configs/single_gpu0_local.yaml \
        scripts/run_orpo.py \
        recipes/scaleup-3stage/orpo-validator-cond-v4.yaml 2>&1 | tee -a "$LOG"
    cd /weka/s225250685/mats-tist
    if [ ! -f "$VAL_COND/config.json" ]; then
        echo "VALIDATOR-COND ORPO FAILED" | tee -a "$LOG"; exit 1
    fi
fi
echo "==== [C] validator-cond ready: $VAL_COND ====" | tee -a "$LOG"

##############################################
# STAGE D+E: Fixer v2 and semantic fixer — reuse from ORPO v2 (already trained)
##############################################
if [ ! -f "$FIXER_V2/config.json" ]; then
    echo "FIXER V2 MISSING at $FIXER_V2" | tee -a "$LOG"; exit 1
fi
echo "==== [D] fixer v2 ready (reusing ORPO v2): $FIXER_V2 ====" | tee -a "$LOG"

if [ ! -f "$SEM_FIXER/config.json" ]; then
    echo "SEMANTIC FIXER MISSING at $SEM_FIXER" | tee -a "$LOG"; exit 1
fi
echo "==== [E] semantic fixer ready (reusing ORPO v2): $SEM_FIXER ====" | tee -a "$LOG"

##############################################
# STAGE F: 5-endpoint rollout (K=8, no-gate sem fixer)
#   H200=141GB: planner(0.30)+val_sel(0.08)+val_cond(0.08)+fixer_v2(0.15)+sem_fixer(0.15)=0.76
##############################################
kill_vllm
echo "==== [F] launching 5 endpoints ====" | tee -a "$LOG"

$VLLM serve "$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"

$VLLM serve "$VAL_SEL" --served-model-name validator_sel --port 8101 --dtype bfloat16 \
    --gpu-memory-utilization 0.08 --enforce-eager --max-model-len 6144 > "${LOG}.vs" 2>&1 &
wait_url http://localhost:8101/v1/models && echo "  validator-sel READY" | tee -a "$LOG"

$VLLM serve "$VAL_COND" --served-model-name validator_cond --port 8104 --dtype bfloat16 \
    --gpu-memory-utilization 0.08 --enforce-eager --max-model-len 6144 > "${LOG}.vc" 2>&1 &
wait_url http://localhost:8104/v1/models && echo "  validator-cond READY" | tee -a "$LOG"

$VLLM serve "$FIXER_V2" --served-model-name fixer --port 8102 --dtype bfloat16 \
    --gpu-memory-utilization 0.15 --enforce-eager --max-model-len 4096 > "${LOG}.f" 2>&1 &
wait_url http://localhost:8102/v1/models && echo "  fixer-v2 READY" | tee -a "$LOG"

$VLLM serve "$SEM_FIXER" --served-model-name fixer_v3 --port 8105 --dtype bfloat16 \
    --gpu-memory-utilization 0.15 --enforce-eager --max-model-len 4096 > "${LOG}.fs" 2>&1 &
wait_url http://localhost:8105/v1/models && echo "  sem-fixer-v3 READY" | tee -a "$LOG"

# sem_fixer_no_gate REMOVED: proven to break 31.7% of correct trajs (10.4x break:rescue).
# Semantic fixer now gated by validator flags only.
OUT=eval_results/scaleup_BoN8_d_K8_3stage_orpov3_sepval_gated_bird_dev.jsonl
rm -f "$OUT"

echo "==== [F] K=8 rollout: exec-err gate + sem-fixer GATED by validator ====" | tee -a "$LOG"
$PY scripts/run_pipeline_rollouts.py \
    --input_file data/sft_bird_with_evidence_dev_text2sql.json \
    --output_file "$OUT" \
    --planner_host http://localhost:8100 \
    --validator_host none \
    --validator_sel_host http://localhost:8101 \
    --validator_cond_host http://localhost:8104 \
    --fixer_host http://localhost:8102 \
    --fixer_gate_exec_ok \
    --fixer_v3_host http://localhost:8105 \
    --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"

echo "==== [F] oracle/greedy metrics ====" | tee -a "$LOG"
$PY scripts/compute_bestofn_metrics.py "$OUT" orpov3_sepval_gated 2>&1 | tee -a "$LOG"

##############################################
# STAGE G: Selector v2
##############################################
kill_vllm
echo "==== [G] selector v2 apply ====" | tee -a "$LOG"
$VLLM serve "$SEL_V2" --served-model-name selector --port 8103 --dtype bfloat16 \
    --gpu-memory-utilization 0.85 --enforce-eager --max-model-len 8192 > "${LOG}.sel" 2>&1 &
wait_url http://localhost:8103/v1/models && echo "  selector READY" | tee -a "$LOG"

label="orpov3_sepval_gated_selectorV2rows"
$PY scripts/compute_bestofn_with_selector.py \
    "$OUT" "$label" --selector_host http://localhost:8103 --row_preview 2>&1 | tee -a "$LOG"

echo "==== ALL_DONE ====" | tee -a "$LOG"