code/slurm_logs/mega_orpo_v2.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=10:00:00
#SBATCH --output=/weka/s225250685/mats-tist/slurm_logs/orpo_v2_%j.out

# ============================================================
# ORPO v2 — improved validators and fixers:
#
# Key improvements over SFT baseline:
#   1. Exec-error fixer: 8203 pairs (4.1x vs 979) using gold SQL fallback
#   2. Semantic fixer: no-gate → runs on ALL exec_ok=True trajectories
#      (trained with 50% preserve pairs to protect correct SQL)
#   3. ORPO for all models: chosen pushed up, rejected pushed down simultaneously
#   4. Validator: ORPO with real exec results + INCORRECT prefix
#
# Oracle ceiling theory:
#   - exec-error fixer recovers 343 questions (22.5%)  → up to 87.9% oracle
#   - semantic fixer (no-gate) recovers 185 more (12.1%) → up to 100% oracle
#   - Realistic target: 73-82% oracle → 65-73% EX with selector
# ============================================================

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
FIXER_V2=$AH/output/fixer-v2-1.5B-execerr-orpo-expanded
VALIDATOR=$AH/output/validator-v4-0.5B-orpo
SEM_FIXER=$AH/output/semantic-fixer-v3-1.5B-orpo-nogate

LOG=/weka/s225250685/mats-tist/slurm_logs/orpo_v2_${SLURM_JOB_ID}.log
: > "$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 all ORPO training data
##############################################
echo "==== [A] building ORPO training data ====" | tee -a "$LOG"

# Fixer v2: 8203 pairs (gold SQL fallback, real exec errors)
# hf_fixer_v2_execerr_expanded is a symlink to hf_fixer_v2_execerr (created once)
echo "  [A1] exec-error fixer data (expanded with gold SQL)..." | tee -a "$LOG"
$PY scripts/build_fixer_v2_execerr.py 2>&1 | tee -a "$LOG"
# Ensure symlink exists (idempotent)
[ -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..." | tee -a "$LOG"
$PY scripts/build_validator_v4_orpo.py 2>&1 | tee -a "$LOG"

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

echo "==== [A] data build done ====" | tee -a "$LOG"
$PY - << 'PYEOF'
from datasets import load_from_disk
for name, path in [("fixer_v2_expanded", "data/hf_fixer_v2_execerr_expanded"),
                    ("validator_v4_orpo", "data/hf_validator_v4_orpo"),
                    ("semantic_fixer_v3", "data/hf_semantic_fixer_v3")]:
    try:
        d = load_from_disk(path)
        split = "train_dpo" if "train_dpo" in d else "train"
        print(f"  {name}: {len(d[split])} train pairs")
    except Exception as e:
        print(f"  {name}: MISSING — {e}")
PYEOF

##############################################
# STAGE B: ORPO — Exec-error Fixer v2 (1.5B)
#   ~8000 pairs, gold SQL + real exec errors
##############################################
echo "==== [B] ORPO fixer v2 exec-error (1.5B, expanded data) ====" | tee -a "$LOG"
cd $AH
PYTHONPATH=src/ ACCELERATE_LOG_LEVEL=info $ACCEL launch \
    --main_process_port 29611 \
    --config_file recipes/accelerate_configs/single_gpu0_local.yaml \
    scripts/run_orpo.py \
    recipes/scaleup-3stage/orpo-fixer-v2-execerr.yaml 2>&1 | tee -a "$LOG"
cd /weka/s225250685/mats-tist

if [ ! -f "$FIXER_V2/config.json" ]; then
    echo "FIXER V2 ORPO FAILED" | tee -a "$LOG"; exit 1
fi
echo "==== [B] fixer v2 saved: $FIXER_V2 ====" | tee -a "$LOG"

##############################################
# STAGE C: ORPO — Validator v4 (0.5B)
#   Binary correct/wrong with exec result in prompt
##############################################
echo "==== [C] ORPO validator v4 (0.5B) ====" | tee -a "$LOG"
cd $AH
PYTHONPATH=src/ ACCELERATE_LOG_LEVEL=info $ACCEL launch \
    --main_process_port 29612 \
    --config_file recipes/accelerate_configs/single_gpu0_local.yaml \
    scripts/run_orpo.py \
    recipes/scaleup-3stage/orpo-validator-v4.yaml 2>&1 | tee -a "$LOG"
cd /weka/s225250685/mats-tist

if [ ! -f "$VALIDATOR/config.json" ]; then
    echo "VALIDATOR ORPO FAILED" | tee -a "$LOG"; exit 1
fi
echo "==== [C] validator v4 saved: $VALIDATOR ====" | tee -a "$LOG"

##############################################
# STAGE D: ORPO — Semantic Fixer v3 (1.5B)
#   No-gate: trained on 50% preserve + 50% wrong→correct
##############################################
echo "==== [D] ORPO semantic fixer v3 (1.5B, no-gate) ====" | tee -a "$LOG"
cd $AH
PYTHONPATH=src/ ACCELERATE_LOG_LEVEL=info $ACCEL launch \
    --main_process_port 29613 \
    --config_file recipes/accelerate_configs/single_gpu0_local.yaml \
    scripts/run_orpo.py \
    recipes/scaleup-3stage/orpo-semantic-fixer-v3.yaml 2>&1 | tee -a "$LOG"
cd /weka/s225250685/mats-tist

if [ ! -f "$SEM_FIXER/config.json" ]; then
    echo "SEMANTIC FIXER ORPO FAILED" | tee -a "$LOG"; exit 1
fi
echo "==== [D] semantic fixer v3 saved: $SEM_FIXER ====" | tee -a "$LOG"

##############################################
# STAGE E: 3-stage K=8 rollout
#   H200=141GB: planner(0.30)+val(0.08+0.08)+fixer_v2(0.15)+sem_fixer(0.15)=0.76
#   Semantic fixer: NO GATE — runs on all exec_ok=True trajectories
#   Exec-error fixer: gated on exec_ok=False
##############################################
kill_vllm
echo "==== [E] 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"

# Validator v4 serves both sel+cond roles (same combined model, different ports)
$VLLM serve "$VALIDATOR" --served-model-name validator_sel --port 8101 --dtype bfloat16 \
    --gpu-memory-utilization 0.08 --enforce-eager --max-model-len 6144 > "${LOG}.vs" 2>&1 &
wait_url http://localhost:8101/v1/models && echo "  validator_v4 sel READY" | tee -a "$LOG"

$VLLM serve "$VALIDATOR" --served-model-name validator_cond --port 8104 --dtype bfloat16 \
    --gpu-memory-utilization 0.08 --enforce-eager --max-model-len 6144 > "${LOG}.vc" 2>&1 &
wait_url http://localhost:8104/v1/models && echo "  validator_v4 cond READY" | tee -a "$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"

OUT=eval_results/scaleup_BoN8_d_K8_3stage_orpov2_nogate_bird_dev.jsonl
rm -f "$OUT"

echo "==== [E] K=8 rollout: fixer_v2 exec-err gate + sem_fixer NO GATE ====" | 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 \
    --sem_fixer_no_gate \
    --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 "==== [E] oracle/greedy metrics ====" | tee -a "$LOG"
$PY scripts/compute_bestofn_metrics.py "$OUT" orpov2_nogate 2>&1 | tee -a "$LOG"

##############################################
# STAGE F: Apply selector v2
##############################################
kill_vllm
echo "==== [F] 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="$(basename $OUT .jsonl)_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"