Add stage1 evaluation reproduction script

scripts/reproduce_stage1_eval.sh

@@ -0,0 +1,60 @@
+#!/usr/bin/env bash
+set -euo pipefail
+
+# Reproduce stage-1 evaluation from a local adapter, local merged model, or HF adapter.
+# Usage examples:
+#   bash scripts/reproduce_stage1_eval.sh runs/qwen3-8b-qlora-20260501-083834/outputs/merged outputs/repro_stage1
+#   bash scripts/reproduce_stage1_eval.sh runs/qwen3-8b-qlora-20260501-083834/outputs/adapter outputs/repro_stage1
+#   bash scripts/reproduce_stage1_eval.sh nraptisss/Qwen3-8B-TMF921-Intent-QLoRA-qwen3-8b-qlora-20260501-083834 outputs/repro_stage1
+
+if [ $# -lt 2 ]; then
+  echo "Usage: $0 <MODEL_OR_ADAPTER_PATH_OR_REPO> <OUTPUT_DIR>" >&2
+  exit 1
+fi
+
+MODEL_OR_ADAPTER="$1"
+OUT_DIR="$2"
+
+source .venv/bin/activate
+export CUDA_VISIBLE_DEVICES="${CUDA_VISIBLE_DEVICES:-0}"
+export PYTHONPATH="$PWD/src:${PYTHONPATH:-}"
+export TOKENIZERS_PARALLELISM=false
+
+python scripts/check_gpu.py
+mkdir -p "$OUT_DIR"
+
+# Local merged model: directory exists and has no adapter_config.json.
+if [ -d "$MODEL_OR_ADAPTER" ] && [ ! -f "$MODEL_OR_ADAPTER/adapter_config.json" ]; then
+  echo "Reproducing eval with local merged model: $MODEL_OR_ADAPTER"
+  python scripts/evaluate_model.py \
+    --model "$MODEL_OR_ADAPTER" \
+    --dataset nraptisss/TMF921-intent-to-config-research-sota \
+    --output_dir "$OUT_DIR" \
+    --batch_size "${EVAL_BATCH_SIZE:-8}" \
+    --max_new_tokens "${EVAL_MAX_NEW_TOKENS:-1536}" \
+    --gold_length_buffer "${EVAL_GOLD_LENGTH_BUFFER:-96}" \
+    --save_every "${EVAL_SAVE_EVERY:-25}"
+else
+  echo "Reproducing eval with PEFT adapter: $MODEL_OR_ADAPTER"
+  python scripts/evaluate_model.py \
+    --model Qwen/Qwen3-8B \
+    --adapter "$MODEL_OR_ADAPTER" \
+    --dataset nraptisss/TMF921-intent-to-config-research-sota \
+    --output_dir "$OUT_DIR" \
+    --load_in_4bit \
+    --batch_size "${EVAL_BATCH_SIZE:-4}" \
+    --max_new_tokens "${EVAL_MAX_NEW_TOKENS:-1536}" \
+    --gold_length_buffer "${EVAL_GOLD_LENGTH_BUFFER:-96}" \
+    --save_every "${EVAL_SAVE_EVERY:-25}"
+fi
+
+python scripts/normalize_eval_metrics.py --eval_dir "$OUT_DIR"
+
+python - <<PY
+import json
+from pathlib import Path
+p = Path("$OUT_DIR") / "all_normalized_metrics.json"
+m = json.loads(p.read_text())
+for split, s in m.items():
+    print(f"{split}: parse={s.get('parse_json'):.4f} norm_field_f1={s.get('norm_field_f1'):.4f} norm_key_f1={s.get('norm_key_f1'):.4f} norm_exact={s.get('norm_exact_match'):.4f}")
+PY