scripts/nohup_eval.sh

#!/usr/bin/env bash
set -euo pipefail

# Evaluate a completed or checkpointed run under nohup.
# Resumable: existing eval/<split>/predictions.json rows are skipped.
# Usage:
#   Adapter mode (default): EVAL_BATCH_SIZE=4 bash scripts/nohup_eval.sh <RUN_DIR> [adapter_path]
#   Merged mode:           EVAL_BATCH_SIZE=8 bash scripts/nohup_eval.sh <RUN_DIR> <merged_model_dir>
#
# If the second path contains adapter_config.json, it is treated as a PEFT adapter.
# Otherwise, it is treated as a merged/full model directory.

if [ $# -lt 1 ]; then
  echo "Usage: $0 <RUN_DIR> [ADAPTER_OR_MERGED_MODEL_PATH]" >&2
  exit 1
fi

RUN_DIR="$1"
MODEL_OR_ADAPTER="${2:-$RUN_DIR/outputs/adapter}"

if [ ! -d "$MODEL_OR_ADAPTER" ]; then
  echo "ERROR: path not found: $MODEL_OR_ADAPTER" >&2
  exit 1
fi

source .venv/bin/activate
export CUDA_VISIBLE_DEVICES="${CUDA_VISIBLE_DEVICES:-0}"
export PYTHONPATH="$PWD/src:${PYTHONPATH:-}"
export TOKENIZERS_PARALLELISM=false

python scripts/check_gpu.py

mkdir -p "$RUN_DIR/logs" "$RUN_DIR/eval"
LOG="$RUN_DIR/logs/eval_$(date +%Y%m%d-%H%M%S).log"

if [ -f "$MODEL_OR_ADAPTER/adapter_config.json" ]; then
  MODE="adapter"
  MODEL_ARG="Qwen/Qwen3-8B"
  ADAPTER_ARGS="--adapter '$MODEL_OR_ADAPTER' --load_in_4bit"
else
  MODE="merged"
  MODEL_ARG="$MODEL_OR_ADAPTER"
  ADAPTER_ARGS=""
fi

nohup bash -lc "
source .venv/bin/activate
export CUDA_VISIBLE_DEVICES='$CUDA_VISIBLE_DEVICES'
export PYTHONPATH='$PWD/src'
export TOKENIZERS_PARALLELISM=false

python scripts/evaluate_model.py \\
  --model '$MODEL_ARG' \\
  $ADAPTER_ARGS \\
  --dataset nraptisss/TMF921-intent-to-config-research-sota \\
  --output_dir '$RUN_DIR/eval' \\
  --batch_size '${EVAL_BATCH_SIZE:-4}' \\
  --max_new_tokens '${EVAL_MAX_NEW_TOKENS:-1536}' \\
  --gold_length_buffer '${EVAL_GOLD_LENGTH_BUFFER:-96}' \\
  --save_every '${EVAL_SAVE_EVERY:-25}'
" > "$LOG" 2>&1 &

echo $! > "$RUN_DIR/EVAL_PID.txt"

cat <<EOF
Started nohup evaluation.
RUN_DIR=$RUN_DIR
MODE=$MODE
MODEL_OR_ADAPTER=$MODEL_OR_ADAPTER
PID=$(cat "$RUN_DIR/EVAL_PID.txt")
LOG=$LOG
RESULTS=$RUN_DIR/eval/all_metrics.json

Monitor:
  tail -f "$LOG"

Tuning knobs:
  EVAL_BATCH_SIZE=${EVAL_BATCH_SIZE:-4}
  EVAL_MAX_NEW_TOKENS=${EVAL_MAX_NEW_TOKENS:-1536}
  EVAL_GOLD_LENGTH_BUFFER=${EVAL_GOLD_LENGTH_BUFFER:-96}
EOF