run_all.sh

#!/usr/bin/env bash
###############################################################################
# StepProbe — Master Orchestrator
# 
# Runs the ENTIRE experiment pipeline from scratch:
#   0. Setup environment
#   1. Download / quantize models
#   2. FP16 baseline inference on all benchmarks
#   3. Quantized inference (all methods x bit-widths x models)
#   4. Segment all CoT traces
#   5. Diagnose step-level errors
#   6. Compute metrics
#   7. Build Silver Bullet datasets + QLoRA restoration
#   8. Re-evaluate restored models
#   9. Generate paper figures + summary tables
#
# Usage:
#   bash run_all.sh                   # full run
#   bash run_all.sh --quick           # 50 samples per benchmark (testing)
#   bash run_all.sh --phase 5         # resume from phase 5
#   bash run_all.sh --models small    # only 1.5B + 7B models
#   bash run_all.sh --dry-run         # print commands without executing
###############################################################################
set -euo pipefail

# Avoid spurious HF download retries on slow/shared bandwidth (default is 10s).
export HF_HUB_DOWNLOAD_TIMEOUT=300

# Pin to a specific GPU (GPU 0 is often shared on this box).
export CUDA_VISIBLE_DEVICES="${CUDA_VISIBLE_DEVICES:-1}"

# ========================== CONFIG ==========================
PROJECT_DIR="$(cd "$(dirname "$0")" && pwd)"
RESULTS_DIR="${PROJECT_DIR}/results"
FIGURES_DIR="${PROJECT_DIR}/figures"
LOGS_DIR="${PROJECT_DIR}/logs"
QUANT_MODELS_DIR="${RESULTS_DIR}/quantized_models"
# llmcompressor and vLLM have mutually exclusive dep pins, so quantization runs
# in its own conda env. Quantized models on disk are env-agnostic.
QUANT_ENV_PYTHON="/home/aiteam1/anaconda3/envs/sonthh-stepprobe-quant/bin/python"
TIMESTAMP=$(date +%Y%m%d_%H%M%S)
LOG_FILE="${LOGS_DIR}/run_all_${TIMESTAMP}.log"

# Defaults
MAX_SAMPLES=""          # empty = full dataset
START_PHASE=0
MODEL_SET="full"        # full | small | primary
DRY_RUN=false
USE_LLM_JUDGE=false
JUDGE_PROVIDER="openai"
SEED=42
NUM_RUNS=1
MAX_TOKENS=4096

# ========================== PARSE ARGS ==========================
while [[ $# -gt 0 ]]; do
    case $1 in
        --quick)        MAX_SAMPLES=50; NUM_RUNS=1; shift ;;
        --medium)       MAX_SAMPLES=200; NUM_RUNS=1; shift ;;
        --phase)        START_PHASE=$2; shift 2 ;;
        --models)       MODEL_SET=$2; shift 2 ;;
        --dry-run)      DRY_RUN=true; shift ;;
        --llm-judge)    USE_LLM_JUDGE=true; shift ;;
        --judge)        JUDGE_PROVIDER=$2; shift 2 ;;
        --samples)      MAX_SAMPLES=$2; shift 2 ;;
        --runs)         NUM_RUNS=$2; shift 2 ;;
        --seed)         SEED=$2; shift 2 ;;
        --help|-h)
            echo "Usage: bash run_all.sh [OPTIONS]"
            echo ""
            echo "Options:"
            echo "  --quick           50 samples per benchmark (fast test)"
            echo "  --medium          200 samples per benchmark"
            echo "  --samples N       Custom sample limit"
            echo "  --phase N         Resume from phase N (0-9)"
            echo "  --models SET      Model set: full|small|primary (default: full)"
            echo "  --runs N          Number of inference runs (default: 1)"
            echo "  --llm-judge       Use LLM judge (costs API money)"
            echo "  --judge PROVIDER  openai|anthropic (default: openai)"
            echo "  --dry-run         Print commands without running"
            echo "  --seed N          Random seed (default: 42)"
            exit 0 ;;
        *) echo "Unknown option: $1"; exit 1 ;;
    esac
done

# ========================== MODEL DEFINITIONS ==========================
# Format: "HF_NAME|TAG|FP16_VRAM_GB|NOTES"
declare -a PRIMARY_MODELS=(
    "deepseek-ai/DeepSeek-R1-Distill-Qwen-7B|r1-qwen-7b|14|primary"
)

declare -a SMALL_MODELS=(
    "deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B|r1-qwen-1.5b|3|small"
    "deepseek-ai/DeepSeek-R1-Distill-Qwen-7B|r1-qwen-7b|14|primary"
)

declare -a FULL_MODELS=(
    "deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B|r1-qwen-1.5b|3|small"
    "deepseek-ai/DeepSeek-R1-Distill-Qwen-7B|r1-qwen-7b|14|primary"
    "deepseek-ai/DeepSeek-R1-Distill-Qwen-14B|r1-qwen-14b|28|4bit-only"
    "deepseek-ai/DeepSeek-R1-Distill-Llama-8B|r1-llama-8b|16|cross-arch"
    "Qwen/Qwen2.5-7B-Instruct|qwen25-7b|14|non-reasoning-control"
)

# Quantization configs: "METHOD|BITS"
# Note: only 4-bit and 8-bit int schemes are supported by vLLM's
# compressed-tensors backend, so w2/w3 GPTQ are omitted. autoawq is 4-bit only.
declare -a QUANT_CONFIGS=(
    "bnb_nf4|4"
    "gptq|4"
    "awq|4"
)

# Benchmarks
declare -a BENCHMARKS=(
    "gsm8k"
    "math500"
    "gpqa"
)

# Select model set
case $MODEL_SET in
    full)    MODELS=("${FULL_MODELS[@]}") ;;
    small)   MODELS=("${SMALL_MODELS[@]}") ;;
    primary) MODELS=("${PRIMARY_MODELS[@]}") ;;
    *)       echo "Unknown model set: $MODEL_SET"; exit 1 ;;
esac

# ========================== HELPERS ==========================
mkdir -p "$LOGS_DIR"

log() {
    local msg="[$(date '+%H:%M:%S')] $1"
    echo "$msg" | tee -a "$LOG_FILE"
}

run_cmd() {
    local cmd="$1"
    if $DRY_RUN; then
        echo "  [DRY-RUN] $cmd"
    else
        log "  CMD: $cmd"
        eval "$cmd" 2>&1 | tee -a "$LOG_FILE"
    fi
}

check_gpu() {
    if command -v nvidia-smi &>/dev/null; then
        nvidia-smi --query-gpu=name,memory.total,memory.used --format=csv,noheader 2>/dev/null || true
    else
        echo "No GPU detected (nvidia-smi not found)"
    fi
}

get_sample_flag() {
    if [[ -n "$MAX_SAMPLES" ]]; then
        echo "--max-samples $MAX_SAMPLES"
    fi
}

can_run_fp16() {
    # Check if model fits in FP16 on 24GB
    local vram=$1
    [[ $vram -le 22 ]]
}

get_quant_for_model() {
    local vram=$1
    if [[ $vram -le 22 ]]; then
        # Can run FP16 + all quantizations
        echo "all"
    else
        # Too big for FP16, only run quantized
        echo "quant-only"
    fi
}

# Return 0 (truthy) only if a PEFT adapter directory contains the files that
# save_pretrained() writes. If training was interrupted the adapter/ subdir
# may not exist at all, or it may exist with only a subset of files written.
# Plain `[[ -d ... ]]` is not enough — use this instead.
is_complete_adapter() {
    local d=$1
    [[ -f "${d}/adapter_config.json" ]] || return 1
    [[ -f "${d}/adapter_model.safetensors" || -f "${d}/adapter_model.bin" ]] || return 1
    return 0
}

# Resolve the model path to pass to run_inference.py:
#   - awq / gptq: local pre-quantized dir (must exist from phase1b)
#   - bnb_nf4:    original HF name (runtime quantization)
#   - fp16:       original HF name
resolve_model_path() {
    local hf_name=$1
    local tag=$2
    local method=$3
    local bits=$4

    if [[ "$method" == "awq" || "$method" == "gptq" ]]; then
        echo "${QUANT_MODELS_DIR}/${tag}_${method}_w${bits}"
    else
        echo "$hf_name"
    fi
}

# ========================== PHASE 0: SETUP ==========================
phase0_setup() {
    log "=============================================="
    log "PHASE 0: Environment Setup"
    log "=============================================="

    log "Project dir: $PROJECT_DIR"
    log "Results dir: $RESULTS_DIR"
    log "Model set:   $MODEL_SET (${#MODELS[@]} models)"
    log "Benchmarks:  ${BENCHMARKS[*]}"
    log "Samples:     ${MAX_SAMPLES:-all}"
    log "Runs:        $NUM_RUNS"
    log "LLM Judge:   $USE_LLM_JUDGE"
    log "GPU:"
    check_gpu | while read -r line; do log "  $line"; done

    # Install dependencies
    log "Installing Python dependencies..."
    run_cmd "pip install -r ${PROJECT_DIR}/requirements.txt --break-system-packages -q"

    # Verify imports
    run_cmd "python -c 'from stepprobe import segment, align, diagnose, metrics, restore, utils; print(\"All modules OK\")'"

    mkdir -p "$RESULTS_DIR"/{inference,segmented,diagnosis,metrics,restored,silver_bullet}
    mkdir -p "$FIGURES_DIR"

    log "Setup complete."
}

# ========================== PHASE 1: DOWNLOAD MODELS ==========================
phase1_download() {
    log "=============================================="
    log "PHASE 1: Download / Verify Models"
    log "=============================================="

    for model_spec in "${MODELS[@]}"; do
        IFS='|' read -r hf_name tag vram notes <<< "$model_spec"
        log "Checking model: $hf_name ($tag)"

        run_cmd "python -c \"
from huggingface_hub import snapshot_download, HfApi
try:
    api = HfApi()
    info = api.model_info('${hf_name}')
    print(f'  Model found: {info.id}, size: {info.siblings and len(info.siblings)} files')
except Exception as e:
    print(f'  Downloading: ${hf_name}...')
    snapshot_download('${hf_name}', local_dir_use_symlinks=True)
\""
    done

    log "Model verification complete."
}

# ========================== PHASE 1b: OFFLINE QUANTIZATION ==========================
# Quantize each (model, method, bits) ONCE with a fixed WikiText-2 calibration
# set. Skipped for bnb_nf4 (runtime quantization) and fp16 (not quantized).
phase1b_quantize() {
    log "=============================================="
    log "PHASE 1b: Offline AWQ / GPTQ Quantization"
    log "=============================================="

    mkdir -p "$QUANT_MODELS_DIR"

    for model_spec in "${MODELS[@]}"; do
        IFS='|' read -r hf_name tag vram notes <<< "$model_spec"

        # Offline quantization loads the full FP16 model with CPU offload.
        # On a 24 GB card this is practical up to ~14B (verified); 32B+ either
        # OOMs or runs for many hours via disk-offloaded forward passes.
        if [[ $vram -gt 30 ]]; then
            log "SKIP quantize for $tag (FP16 size ${vram}GB too large for offline quantization on 24GB GPU)"
            continue
        fi

        for quant_spec in "${QUANT_CONFIGS[@]}"; do
            IFS='|' read -r method bits <<< "$quant_spec"

            # Skip methods that don't need offline quantization
            [[ "$method" == "awq" || "$method" == "gptq" ]] || continue

            local out_dir="${QUANT_MODELS_DIR}/${tag}_${method}_w${bits}"

            if [[ -f "${out_dir}/config.json" ]]; then
                log "SKIP (exists): $out_dir"
                continue
            fi

            log "Quantizing: $tag / $method / w${bits}"
            run_cmd "$QUANT_ENV_PYTHON ${PROJECT_DIR}/scripts/quantize_models.py \
                --model $hf_name \
                --method $method \
                --bits $bits \
                --group-size 128 \
                --output $out_dir"

            # Free GPU between quantization jobs
            run_cmd "$QUANT_ENV_PYTHON -c 'import torch; torch.cuda.empty_cache() if torch.cuda.is_available() else None'"
        done
    done

    log "Quantization complete. Models in: $QUANT_MODELS_DIR"
}

# ========================== PHASE 2: FP16 BASELINE ==========================
phase2_fp16_inference() {
    log "=============================================="
    log "PHASE 2: FP16 Baseline Inference"
    log "=============================================="

    local sample_flag=$(get_sample_flag)

    for model_spec in "${MODELS[@]}"; do
        IFS='|' read -r hf_name tag vram notes <<< "$model_spec"

        if ! can_run_fp16 "$vram"; then
            log "SKIP FP16 for $tag (needs ${vram}GB > 24GB VRAM)"
            continue
        fi

        for bench in "${BENCHMARKS[@]}"; do
            local out_dir="${RESULTS_DIR}/inference/fp16/${tag}"
            local out_file="${out_dir}/${bench}_run0.jsonl"

            if [[ -f "$out_file" ]]; then
                log "SKIP (exists): $out_file"
                continue
            fi

            log "Running FP16 inference: $tag / $bench"
            for run_idx in $(seq 0 $((NUM_RUNS - 1))); do
                run_cmd "python ${PROJECT_DIR}/scripts/run_inference.py \
                    --model $hf_name \
                    --benchmark $bench \
                    --output $out_dir \
                    --max-tokens $MAX_TOKENS \
                    --num-runs 1 \
                    $sample_flag"
            done
        done

        # Free GPU memory between models
        run_cmd "python -c 'import torch; torch.cuda.empty_cache() if torch.cuda.is_available() else None'"
    done
}

# ========================== PHASE 3: QUANTIZED INFERENCE ==========================
phase3_quantized_inference() {
    log "=============================================="
    log "PHASE 3: Quantized Inference"
    log "=============================================="

    local sample_flag=$(get_sample_flag)

    for model_spec in "${MODELS[@]}"; do
        IFS='|' read -r hf_name tag vram notes <<< "$model_spec"

        for quant_spec in "${QUANT_CONFIGS[@]}"; do
            IFS='|' read -r method bits <<< "$quant_spec"
            local quant_tag="${method}_w${bits}"

            # Skip 3-bit and 2-bit for very large models (too slow / unstable)
            if [[ $vram -ge 28 && $bits -lt 4 ]]; then
                log "SKIP $quant_tag for $tag (large model + low bit)"
                continue
            fi

            for bench in "${BENCHMARKS[@]}"; do
                local out_dir="${RESULTS_DIR}/inference/${quant_tag}/${tag}"
                local out_file="${out_dir}/${bench}_run0.jsonl"

                if [[ -f "$out_file" ]]; then
                    log "SKIP (exists): $out_file"
                    continue
                fi

                local model_path=$(resolve_model_path "$hf_name" "$tag" "$method" "$bits")

                # For awq/gptq, the local dir must exist (from phase1b)
                if [[ "$method" == "awq" || "$method" == "gptq" ]] && [[ ! -f "${model_path}/config.json" ]]; then
                    log "SKIP $quant_tag / $tag / $bench: quantized model not found at $model_path"
                    continue
                fi

                log "Running $quant_tag inference: $tag / $bench (from $model_path)"
                run_cmd "python ${PROJECT_DIR}/scripts/run_inference.py \
                    --model $model_path \
                    --quant $method \
                    --bits $bits \
                    --benchmark $bench \
                    --output $out_dir \
                    --max-tokens $MAX_TOKENS \
                    --num-runs 1 \
                    $sample_flag"
            done

            # Free GPU
            run_cmd "python -c 'import torch; torch.cuda.empty_cache() if torch.cuda.is_available() else None'"
        done
    done
}

# ========================== PHASE 4: SEGMENTATION ==========================
phase4_segment() {
    log "=============================================="
    log "PHASE 4: CoT Step Segmentation"
    log "=============================================="

    # Segment all inference outputs
    for inf_dir in "${RESULTS_DIR}"/inference/*/; do
        local quant_tag=$(basename "$inf_dir")

        for model_dir in "${inf_dir}"*/; do
            [[ -d "$model_dir" ]] || continue
            local model_tag=$(basename "$model_dir")
            local seg_dir="${RESULTS_DIR}/segmented/${quant_tag}/${model_tag}"

            for jsonl_file in "${model_dir}"*.jsonl; do
                [[ -f "$jsonl_file" ]] || continue
                local basename_f=$(basename "$jsonl_file")
                local out_file="${seg_dir}/${basename_f}"

                if [[ -f "$out_file" ]]; then
                    log "SKIP (exists): $out_file"
                    continue
                fi

                log "Segmenting: ${quant_tag}/${model_tag}/${basename_f}"
                mkdir -p "$seg_dir"
                run_cmd "python -m stepprobe.segment \
                    --input $(dirname $jsonl_file) \
                    --output $seg_dir \
                    --model $model_tag \
                    --quant $quant_tag"
            done
        done
    done
}

# ========================== PHASE 5: DIAGNOSIS ==========================
phase5_diagnose() {
    log "=============================================="
    log "PHASE 5: Step-Level Error Diagnosis"
    log "=============================================="

    local judge_flags=""
    if $USE_LLM_JUDGE; then
        judge_flags="--judge $JUDGE_PROVIDER --judge-model gpt-4o"
    fi

    for model_spec in "${MODELS[@]}"; do
        IFS='|' read -r hf_name tag vram notes <<< "$model_spec"

        # Find FP16 reference (use own FP16 if available, else skip)
        local ref_dir="${RESULTS_DIR}/segmented/fp16/${tag}"
        if [[ ! -d "$ref_dir" ]]; then
            log "WARN: No FP16 reference for $tag. Using closest available."
            # For models too large for FP16, use 8-bit as reference
            ref_dir="${RESULTS_DIR}/segmented/bnb_nf4/${tag}"
            if [[ ! -d "$ref_dir" ]]; then
                log "SKIP diagnosis for $tag (no reference traces)"
                continue
            fi
        fi

        for quant_spec in "${QUANT_CONFIGS[@]}"; do
            IFS='|' read -r method bits <<< "$quant_spec"
            local quant_tag="${method}_w${bits}"
            local hyp_dir="${RESULTS_DIR}/segmented/${quant_tag}/${tag}"

            [[ -d "$hyp_dir" ]] || continue

            local diag_dir="${RESULTS_DIR}/diagnosis/${quant_tag}/${tag}"

            # Check if already done
            local any_missing=false
            for jsonl_file in "${hyp_dir}"/*.jsonl; do
                [[ -f "$jsonl_file" ]] || continue
                local basename_f=$(basename "$jsonl_file")
                [[ -f "${diag_dir}/${basename_f}" ]] || any_missing=true
            done

            if ! $any_missing && [[ -d "$diag_dir" ]]; then
                log "SKIP (exists): diagnosis for ${quant_tag}/${tag}"
                continue
            fi

            log "Diagnosing: ${quant_tag} / ${tag}"
            mkdir -p "$diag_dir"
            run_cmd "python -m stepprobe.diagnose \
                --ref $ref_dir \
                --hyp $hyp_dir \
                --output $diag_dir \
                --alignment dtw \
                $judge_flags"
        done
    done
}

# ========================== PHASE 6: METRICS ==========================
phase6_metrics() {
    log "=============================================="
    log "PHASE 6: Compute StepProbe Metrics"
    log "=============================================="

    local metrics_dir="${RESULTS_DIR}/metrics"
    mkdir -p "$metrics_dir"

    for model_spec in "${MODELS[@]}"; do
        IFS='|' read -r hf_name tag vram notes <<< "$model_spec"

        # Compute FP16 accuracy first (for deltas)
        local fp16_acc=""
        local fp16_diag="${RESULTS_DIR}/diagnosis/fp16/${tag}"
        # (FP16 doesn't have diagnosis, compute from inference)

        for quant_spec in "${QUANT_CONFIGS[@]}"; do
            IFS='|' read -r method bits <<< "$quant_spec"
            local quant_tag="${method}_w${bits}"
            local diag_dir="${RESULTS_DIR}/diagnosis/${quant_tag}/${tag}"

            [[ -d "$diag_dir" ]] || continue

            local out_prefix="${metrics_dir}/${tag}_${quant_tag}"

            log "Computing metrics: ${tag} / ${quant_tag}"
            run_cmd "python -m stepprobe.metrics \
                --diagnosis $diag_dir \
                --output $metrics_dir \
                --model $tag \
                --quant $quant_tag"
        done
    done
}

# ========================== PHASE 7: RESTORATION ==========================
phase7_restore() {
    log "=============================================="
    log "PHASE 7: Targeted Restoration (QLoRA + DPO)"
    log "=============================================="

    for model_spec in "${MODELS[@]}"; do
        IFS='|' read -r hf_name tag vram notes <<< "$model_spec"

        # Only restore models that fit for QLoRA (need FP16 ref + 4bit base)
        if [[ $vram -gt 16 ]]; then
            log "SKIP restoration for $tag (too large for QLoRA on 24GB)"
            continue
        fi

        local ref_dir="${RESULTS_DIR}/segmented/fp16/${tag}"
        [[ -d "$ref_dir" ]] || continue

        # Restore for each quantization method
        for quant_spec in "${QUANT_CONFIGS[@]}"; do
            IFS='|' read -r method bits <<< "$quant_spec"
            local quant_tag="${method}_w${bits}"
            local diag_dir="${RESULTS_DIR}/diagnosis/${quant_tag}/${tag}"

            [[ -d "$diag_dir" ]] || continue

            local restore_dir="${RESULTS_DIR}/restored/${quant_tag}/${tag}"

            if is_complete_adapter "${restore_dir}/qlora/adapter"; then
                log "SKIP (exists): QLoRA restoration for ${quant_tag}/${tag}"
            else
                # Partial state from an interrupted run? HF Trainer will
                # overwrite, but make it loud so stale checkpoints aren't mistaken for success.
                if [[ -d "${restore_dir}/qlora" ]]; then
                    log "WARN: partial QLoRA state found at ${restore_dir}/qlora — restarting training (will overwrite)"
                fi
                log "QLoRA restoration: ${tag} / ${quant_tag}"
                run_cmd "python -m stepprobe.restore \
                    --model $hf_name \
                    --diagnosis $diag_dir \
                    --ref $ref_dir \
                    --output $restore_dir \
                    --method qlora \
                    --max-samples 500 \
                    --epochs 3 \
                    --lr 2e-4 \
                    --batch-size 4"
            fi

            # Also try DPO. NOTE: we write DPO output to "${restore_dir}_dpo"
            # (suffixed tag dir) so QLoRA and DPO don't collide — the skip
            # check must therefore look at "${restore_dir}_dpo/dpo/adapter",
            # not "${restore_dir}/dpo/adapter".
            if is_complete_adapter "${restore_dir}_dpo/dpo/adapter"; then
                log "SKIP (exists): DPO restoration for ${quant_tag}/${tag}"
            else
                if [[ -d "${restore_dir}_dpo/dpo" ]]; then
                    log "WARN: partial DPO state found at ${restore_dir}_dpo/dpo — restarting training (will overwrite)"
                fi
                log "DPO restoration: ${tag} / ${quant_tag}"
                run_cmd "python -m stepprobe.restore \
                    --model $hf_name \
                    --diagnosis $diag_dir \
                    --ref $ref_dir \
                    --output ${restore_dir}_dpo \
                    --method dpo \
                    --max-samples 300 \
                    --epochs 1 \
                    --lr 5e-5 \
                    --batch-size 2"
            fi

            # Free GPU
            run_cmd "python -c 'import torch; torch.cuda.empty_cache() if torch.cuda.is_available() else None'"
        done
    done
}

# ========================== PHASE 8: RE-EVALUATE RESTORED ==========================
phase8_reeval() {
    log "=============================================="
    log "PHASE 8: Re-evaluate Restored Models"
    log "=============================================="

    local sample_flag=$(get_sample_flag)

    # Fast restored-inference path:
    #   1. Merge the LoRA adapter into an FP16 copy of the base model (one-time
    #      per (model, quant_tag)). Merging in FP16 is lossless; merging into
    #      4-bit weights — as the old path did — triggers PEFT's rounding-error
    #      warning.
    #   2. Run vLLM on the merged model with --quant bnb_nf4 so it re-quantizes
    #      to NF4 at load. Same deployment target as before, with vLLM's batched
    #      generation instead of HF's one-sample-at-a-time loop.
    #   3. Delete the merged-FP16 dir once all benchmarks for that config are
    #      done, so disk usage stays bounded to ~one model at a time.
    for model_spec in "${MODELS[@]}"; do
        IFS='|' read -r hf_name tag vram notes <<< "$model_spec"

        for quant_spec in "${QUANT_CONFIGS[@]}"; do
            IFS='|' read -r method bits <<< "$quant_spec"
            local quant_tag="${method}_w${bits}"
            local adapter_dir="${RESULTS_DIR}/restored/${quant_tag}/${tag}/qlora/adapter"
            local merged_dir="${RESULTS_DIR}/restored/${quant_tag}/${tag}/qlora/merged_fp16"

            [[ -d "$adapter_dir" ]] || continue

            # Determine which benchmarks still need to run for this config.
            local pending_benches=()
            for bench in "${BENCHMARKS[@]}"; do
                local out_file="${RESULTS_DIR}/inference/${quant_tag}_restored/${tag}/${bench}_run0.jsonl"
                if [[ -f "$out_file" ]]; then
                    log "SKIP (exists): $out_file"
                else
                    pending_benches+=("$bench")
                fi
            done

            if [[ ${#pending_benches[@]} -eq 0 ]]; then
                continue
            fi

            # Merge adapter → FP16 once (cached on disk if re-run).
            if [[ ! -f "${merged_dir}/config.json" ]]; then
                log "Merging adapter into FP16: ${tag} / ${quant_tag}"
                run_cmd "python ${PROJECT_DIR}/scripts/merge_adapter.py \
                    --model $hf_name \
                    --adapter $adapter_dir \
                    --output $merged_dir"
                run_cmd "python -c 'import torch; torch.cuda.empty_cache() if torch.cuda.is_available() else None'"
            fi

            for bench in "${pending_benches[@]}"; do
                local out_dir="${RESULTS_DIR}/inference/${quant_tag}_restored/${tag}"
                mkdir -p "$out_dir"

                log "Re-evaluating restored (vLLM+NF4): ${tag} / ${quant_tag} / ${bench}"
                run_cmd "python ${PROJECT_DIR}/scripts/run_inference.py \
                    --model $merged_dir \
                    --quant bnb_nf4 \
                    --bits $bits \
                    --benchmark $bench \
                    --output $out_dir \
                    --max-tokens $MAX_TOKENS \
                    --num-runs 1 \
                    $sample_flag"
            done

            # Disk hygiene: the merged FP16 dir is ~3-16 GB and is only needed
            # during inference. Once all 3 benchmarks are done, drop it.
            if [[ -d "$merged_dir" ]] && ! $DRY_RUN; then
                log "Cleaning up merged FP16 dir: $merged_dir"
                rm -rf "$merged_dir"
            fi

            run_cmd "python -c 'import torch; torch.cuda.empty_cache() if torch.cuda.is_available() else None'"
        done
    done

    # Re-run segment + diagnose + metrics on restored outputs. `stepprobe.segment`
    # reads every jsonl in the input dir in one invocation, so we call it once
    # per (quant_tag, model_tag) — not once per jsonl.
    log "Segmenting + diagnosing restored model outputs..."
    shopt -s nullglob
    for inf_dir in "${RESULTS_DIR}"/inference/*_restored/; do
        [[ -d "$inf_dir" ]] || continue
        local quant_tag=$(basename "$inf_dir")

        for model_dir in "${inf_dir}"*/; do
            [[ -d "$model_dir" ]] || continue
            local model_tag=$(basename "$model_dir")
            local jsonls=("${model_dir}"*.jsonl)
            [[ ${#jsonls[@]} -gt 0 ]] || continue

            local seg_dir="${RESULTS_DIR}/segmented/${quant_tag}/${model_tag}"
            mkdir -p "$seg_dir"
            run_cmd "python -m stepprobe.segment --input $model_dir --output $seg_dir --quant ${quant_tag}"

            # Diagnose vs FP16 reference.
            local ref_dir="${RESULTS_DIR}/segmented/fp16/${model_tag}"
            [[ -d "$ref_dir" ]] || continue
            local diag_dir="${RESULTS_DIR}/diagnosis/${quant_tag}/${model_tag}"
            mkdir -p "$diag_dir"
            run_cmd "python -m stepprobe.diagnose --ref $ref_dir --hyp $seg_dir --output $diag_dir --alignment dtw"

            # Metrics — tagged so per-(model, quant) files don't clobber.
            run_cmd "python -m stepprobe.metrics --diagnosis $diag_dir --output ${RESULTS_DIR}/metrics --model $model_tag --quant $quant_tag"
        done
    done
    shopt -u nullglob
}

# ========================== PHASE 9: FIGURES & SUMMARY ==========================
phase9_figures() {
    log "=============================================="
    log "PHASE 9: Generate Figures & Summary"
    log "=============================================="

    # Exploratory per-(model, benchmark) figures — useful while iterating;
    # end up as supplementary material in the paper.
    log "Generating exploratory figures (per model × benchmark)..."
    run_cmd "python ${PROJECT_DIR}/scripts/make_figures.py \
        --metrics ${RESULTS_DIR}/metrics \
        --output $FIGURES_DIR"

    # Bootstrap confidence intervals and paired significance tests on the
    # per-problem diagnosis data. Writes *_ci.json and *_sig.json alongside
    # the *_metrics.json files; paper figures consume them for CI bands and
    # significance stars. Needs no GPU — pure numpy resampling.
    log "Computing bootstrap CIs + paired significance tests..."
    run_cmd "python ${PROJECT_DIR}/scripts/compute_ci.py \
        --diagnosis ${RESULTS_DIR}/diagnosis \
        --output ${RESULTS_DIR}/metrics \
        --n-boot 5000"

    # Paper-ready headline figures + LaTeX numerics table.
    # Produces 6 figures: pipeline schematic, SSR (with CI bands), error mix,
    # forest plot with CIs + sig stars, FFS distribution, qualitative trace
    # example. Reads metrics/ (point estimates + CIs + sig) plus diagnosis/
    # and segmented/ (for the per-problem data behind figs 4 and 5).
    log "Generating paper figures + LaTeX table..."
    run_cmd "python ${PROJECT_DIR}/scripts/make_paper_figures.py \
        --metrics ${RESULTS_DIR}/metrics \
        --diagnosis ${RESULTS_DIR}/diagnosis \
        --segmented ${RESULTS_DIR}/segmented \
        --output ${FIGURES_DIR}/paper"

    # Auto-populated supplementary tables (ablation / baselines / interventions).
    # Safe to call even before those experiments have run — placeholders fill
    # in and the paper still compiles.
    log "Generating supplementary LaTeX tables..."
    run_cmd "python ${PROJECT_DIR}/scripts/make_tables.py \
        --metrics-dir ${RESULTS_DIR}/metrics \
        --output-dir ${FIGURES_DIR}/paper"

    # Terminal summary table (for the log).
    log "Generating summary table..."
    run_cmd "python ${PROJECT_DIR}/scripts/summary_table.py --metrics ${RESULTS_DIR}/metrics"

    log "All figures saved to: $FIGURES_DIR"
}

# ========================== MAIN ==========================
main() {
    log "=============================================="
    log "StepProbe — Full Experiment Pipeline"
    log "Started: $(date)"
    log "Config: models=$MODEL_SET, samples=${MAX_SAMPLES:-all}, runs=$NUM_RUNS"
    log "=============================================="

    cd "$PROJECT_DIR"

    [[ $START_PHASE -le 0 ]] && phase0_setup
    [[ $START_PHASE -le 1 ]] && phase1_download
    [[ $START_PHASE -le 1 ]] && phase1b_quantize
    [[ $START_PHASE -le 2 ]] && phase2_fp16_inference
    [[ $START_PHASE -le 3 ]] && phase3_quantized_inference
    [[ $START_PHASE -le 4 ]] && phase4_segment
    [[ $START_PHASE -le 5 ]] && phase5_diagnose
    [[ $START_PHASE -le 6 ]] && phase6_metrics
    [[ $START_PHASE -le 7 ]] && phase7_restore
    [[ $START_PHASE -le 8 ]] && phase8_reeval
    [[ $START_PHASE -le 9 ]] && phase9_figures

    log ""
    log "=============================================="
    log "ALL PHASES COMPLETE"
    log "Finished: $(date)"
    log "Results:  $RESULTS_DIR"
    log "Figures:  $FIGURES_DIR"
    log "Log:      $LOG_FILE"
    log "=============================================="
}

main