run_all_models.sh

#!/bin/bash
# ============================================================================
#  Run Drift Disentanglement on All 5 Models
#  ============================================================================
#  Usage:
#    chmod +x run_all_models.sh
#    ./run_all_models.sh              # Run all models
#    ./run_all_models.sh llama2       # Run single model
#    ./run_all_models.sh --skip_qwen  # Skip Qwen (already done)
#
#  For background execution:
#    tmux new-session -d -s models './run_all_models.sh --skip_qwen 2>&1 | tee run_all.log'
#    tmux attach -t models           # Check progress
# ============================================================================

set -e
export HF_TOKEN="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"

BASE_DIR="$HOME/sv
_kg/knowledge_drift"
DATA_DIR="$BASE_DIR/data"
SCRIPT="$BASE_DIR/disentanglement_v2.py"
GPU="${CUDA_VISIBLE_DEVICES:-0}"

# Model configurations: key|hf_path|dataset|output_dir
declare -A MODEL_HF=(
    ["llama2"]="meta-llama/Llama-2-7b-chat-hf"
    ["mistral"]="mistralai/Mistral-7B-Instruct-v0.3"
    ["llama31"]="meta-llama/Llama-3.1-8B-Instruct"
    ["qwen25"]="Qwen/Qwen2.5-7B-Instruct"
    ["gemma2"]="google/gemma-2-9b-it"
)

# Order: run models from largest dataset (most drifted) to smallest
# LLaMA-2 has ~5265 drifted, Mistral/LLaMA-3.1/Qwen ~2313, Gemma ~1178
RUN_ORDER=("llama2" "mistral" "llama31" "qwen25" "gemma2")

# ============================================================================
#  Parse arguments
# ============================================================================
SKIP_QWEN=false
SKIP_EXTRACTION=false
SINGLE_MODEL=""

for arg in "$@"; do
    case $arg in
        --skip_qwen)    SKIP_QWEN=true ;;
        --skip_extract) SKIP_EXTRACTION=true ;;
        llama2|mistral|llama31|qwen25|gemma2)
            SINGLE_MODEL="$arg" ;;
        *)
            echo "Unknown arg: $arg"
            echo "Usage: $0 [--skip_qwen] [--skip_extract] [model_name]"
            exit 1 ;;
    esac
done

# ============================================================================
#  Preflight checks
# ============================================================================
echo "============================================================"
echo "  MULTI-MODEL DRIFT DETECTION PIPELINE"
echo "============================================================"
echo "  Base dir:  $BASE_DIR"
echo "  GPU:       $GPU"
echo "  Script:    $SCRIPT"
echo "  Skip Qwen: $SKIP_QWEN"
echo ""

# Check datasets exist
MISSING=0
for key in "${RUN_ORDER[@]}"; do
    ds="$DATA_DIR/tier1_${key}.json"
    if [ ! -f "$ds" ]; then
        echo "  ❌ Missing dataset: $ds"
        MISSING=$((MISSING + 1))
    else
        n=$(python3 -c "import json; d=json.load(open('$ds')); s=d.get('samples',d); print(len(s))" 2>/dev/null || echo "?")
        echo "  ✅ $ds ($n samples)"
    fi
done

if [ $MISSING -gt 0 ]; then
    echo ""
    echo "  Missing $MISSING datasets. Run prepare_all_models.py first:"
    echo "    python prepare_all_models.py --master_dataset data/tier1.json"
    echo ""
    # Don't exit — maybe they only want to run existing ones
fi

echo "============================================================"
echo ""

# ============================================================================
#  Run function
# ============================================================================
run_model() {
    local key=$1
    local hf_path="${MODEL_HF[$key]}"
    local dataset="$DATA_DIR/tier1_${key}.json"
    local output_dir="$DATA_DIR/experiments/tier1_${key}_v2"
    local cache="$output_dir/cached_states.npz"

    echo ""
    echo "╔══════════════════════════════════════════════════════════╗"
    echo "║  MODEL: $hf_path"
    echo "║  Dataset: $dataset"
    echo "║  Output:  $output_dir"
    echo "╚══════════════════════════════════════════════════════════╝"
    echo ""

    # Check if dataset exists
    if [ ! -f "$dataset" ]; then
        echo "  ⚠️  Dataset not found: $dataset — SKIPPING"
        return 1
    fi

    # Build command
    local CMD="CUDA_VISIBLE_DEVICES=$GPU python $SCRIPT"
    CMD="$CMD --model $hf_path"
    CMD="$CMD --dataset $dataset"
    CMD="$CMD --output_dir $output_dir"

    # Skip extraction if cache exists or flag set
    if [ -f "$cache" ] || [ "$SKIP_EXTRACTION" = true ]; then
        if [ -f "$cache" ]; then
            echo "  📦 Cache found: $cache — skipping extraction"
        fi
        CMD="$CMD --skip_extraction"
    fi

    echo "  ▶ Running: $CMD"
    echo "  ▶ Started: $(date)"
    local START=$SECONDS

    # Run
    eval $CMD

    local ELAPSED=$((SECONDS - START))
    local MINS=$((ELAPSED / 60))
    local SECS=$((ELAPSED % 60))
    echo ""
    echo "  ✅ $key complete in ${MINS}m ${SECS}s"
    echo "  📊 Results: $output_dir/final_results.json"
    echo ""
}

# ============================================================================
#  Execute
# ============================================================================
TOTAL_START=$SECONDS
COMPLETED=0
FAILED=0

if [ -n "$SINGLE_MODEL" ]; then
    # Run single model
    run_model "$SINGLE_MODEL" && COMPLETED=$((COMPLETED+1)) || FAILED=$((FAILED+1))
else
    # Run all models in order
    for key in "${RUN_ORDER[@]}"; do
        if [ "$SKIP_QWEN" = true ] && [ "$key" = "qwen25" ]; then
            echo "  ⏭  Skipping Qwen2.5 (already done)"
            continue
        fi
        run_model "$key" && COMPLETED=$((COMPLETED+1)) || FAILED=$((FAILED+1))
    done
fi

TOTAL_ELAPSED=$((SECONDS - TOTAL_START))
TOTAL_MINS=$((TOTAL_ELAPSED / 60))
TOTAL_HOURS=$((TOTAL_MINS / 60))
REM_MINS=$((TOTAL_MINS % 60))

echo ""
echo "╔══════════════════════════════════════════════════════════╗"
echo "║  ALL DONE                                               ║"
echo "║  Completed: $COMPLETED  |  Failed: $FAILED                            ║"
echo "║  Total time: ${TOTAL_HOURS}h ${REM_MINS}m                                    ║"
echo "╚══════════════════════════════════════════════════════════╝"
echo ""

# ============================================================================
#  Post-run: Collect results summary
# ============================================================================
echo "RESULTS SUMMARY:"
echo "─────────────────────────────────────────────────────────"
printf "  %-25s %-8s %-8s %-10s %-10s\n" "Model" "Drift" "Unc" "cos(d,u)" "Best Layer"
echo "─────────────────────────────────────────────────────────"

for key in "${RUN_ORDER[@]}"; do
    rf="$DATA_DIR/experiments/tier1_${key}_v2/final_results.json"
    if [ -f "$rf" ]; then
        python3 -c "
import json
with open('$rf') as f: r = json.load(f)
bl = r['best_layer']
lr = r['layer_results'][str(bl)]
print(f\"  ${MODEL_HF[$key]:<25s}  {lr['drift_auroc']:.4f}  {lr['uncertainty_auroc']:.4f}  {lr['cos_drift_uncertainty']:+.4f}     {bl}\")
" 2>/dev/null || echo "  ${MODEL_HF[$key]}: (parse error)"
    else
        echo "  ${MODEL_HF[$key]}: (no results)"
    fi
done
echo "─────────────────────────────────────────────────────────"