variant_a/scripts/run_eval.sh

#!/bin/bash
set -euo pipefail
export PATH="/venv/main/bin:$PATH"
export PYTHONDONTWRITEBYTECODE=1
export CUDA_VISIBLE_DEVICES=0
export WORK=/dev/shm/eval
export QUIP_DIR="$WORK/quip-sharp"

log() { echo "[$(date '+%Y-%m-%d %H:%M:%S')] $*"; }

log "============================================"
log "  Polish LLM Eval — QuIP# Bielik Variant A"
log "============================================"

# Phase 1: Setup
log "Phase 1: Running setup..."
bash /dev/shm/eval/setup_eval.sh 2>&1
log "Setup done"

# Phase 2: Run evaluation
log "Phase 2: Running Polish eval (5-shot)..."

cd "$WORK/lm-evaluation-harness"

python "$WORK/eval_polish_quip.py" \
    --model_path "$WORK/model" \
    --tokenizer "speakleash/Bielik-11B-v2.3-Instruct" \
    --output_dir "$WORK/results_a" \
    --num_fewshot 5 \
    --batch_size 1 \
    --tasks polish_generate_few polish_mc \
    2>&1 | tee "$WORK/eval_log.txt"

log "============================================"
log "  Evaluation complete!"
log "  Results: $WORK/results_a/"
log "  Log:     $WORK/eval_log.txt"
log "============================================"