step_train/scripts_train/run_eval_splits.sh

#!/bin/bash
# 跑完 SFT/DAPO 后的全流程评测 —— 切分消融四套模型(by-query 1:2 / by-user 2:1 / by-user 1:2)
# × {LoRA, Full} × {SFT warm-start, DAPO 终产物}。
#
# 完全复用既有评测口径(scripts/train/11_eval.py):greedy temp=0、复用 RL rollout、
# BGE-M3 user-turn-only 检索、预算 800、gpt-4o-mini judge、按 MS/query_type 聚合。
# 评测集 = data/processed/metamem_5k/dataset_eval100.json(固定 100 用户 / 308 QA,seed=42,
# 所有模型共用同一份 → 公平对照),与既有 eval_metamem_trainset_qwen25_sft_{lora,full} 同集。
#
# ⚠️ 设计与既有约定一致:
#   - 每个被评 checkpoint 一个 config(configs/eval/eval_trainset_qwen25_*),输出各自独立
#     data/eval/<name>.jsonl + .stats.json,互不覆盖。
#   - 11_eval.py 自带 resume(读已有 jsonl 的 query_id 跳过)→ 崩了重跑即续。
#   - checkpoint 不存在(还没训完)的 config 自动跳过,不报错 → 可边训边评。
#   - judge 缓存按「答案文本」key,跨 checkpoint 复用安全(同答案→同判定),省 API。
#
# 用法:
#   bash scripts/train/run_eval_splits.sh                 # 评所有已就绪的 checkpoint
#   bash scripts/train/run_eval_splits.sh --only dapo     # 只评 DAPO 终产物
#   bash scripts/train/run_eval_splits.sh --only sft      # 只评 SFT warm-start
#   ONLY_TAG=bu_2to1 bash scripts/train/run_eval_splits.sh # 只评某一套切分
#   NO_JUDGE=1 bash scripts/train/run_eval_splits.sh       # 离线:跳过 gpt-4o-mini(只出 EM/F1+label 分布)
set -u

PROJECT_ROOT=/mnt/train-gui-agent/zhangzeyu/mem_memory
cd "$PROJECT_ROOT"
PY=${PYTHON:-/root/mambaforge/envs/gjh_memo/bin/python}

ONLY=""          # sft | dapo | ""(both)
[ "${1:-}" = "--only" ] && ONLY="${2:-}"
ONLY_TAG="${ONLY_TAG:-}"   # bu_2to1 | bu_1to2 | 1to2 | ""(all)
JUDGE_FLAG=""
[ "${NO_JUDGE:-0}" = "1" ] && JUDGE_FLAG="--no-judge"

# 每行: <stage> <split_tag> <eval_config> <checkpoint_dir_that_must_exist>
# checkpoint_dir 是 11_eval 真正加载权重的目录(LoRA→含 capability_lora.safetensors;
# Full→含 model-*.safetensors)。DAPO 的终产物在 <output_dir>/final。
ROWS=$(cat <<'EOF'
sft  bu_2to1 configs/eval/eval_trainset_qwen25_sft_lora_bu_2to1.yaml  checkpoints/sft/Qwen2.5-7B-Instruct_bu_2to1/capability_lora.safetensors
sft  bu_1to2 configs/eval/eval_trainset_qwen25_sft_lora_bu_1to2.yaml  checkpoints/sft/Qwen2.5-7B-Instruct_bu_1to2/capability_lora.safetensors
sft  1to2    configs/eval/eval_trainset_qwen25_sft_lora_1to2.yaml     checkpoints/sft/Qwen2.5-7B-Instruct_1to2/capability_lora.safetensors
sft  bu_2to1 configs/eval/eval_trainset_qwen25_sft_full_bu_2to1.yaml  checkpoints/sft_full/Qwen2.5-7B-Instruct_bu_2to1/done.flag
sft  bu_1to2 configs/eval/eval_trainset_qwen25_sft_full_bu_1to2.yaml  checkpoints/sft_full/Qwen2.5-7B-Instruct_bu_1to2/done.flag
sft  1to2    configs/eval/eval_trainset_qwen25_sft_full_1to2.yaml     checkpoints/sft_full/Qwen2.5-7B-Instruct_1to2/done.flag
dapo bu_2to1 configs/eval/eval_trainset_qwen25_dapo_lora_bu_2to1.yaml checkpoints/dapo/Qwen2.5-7B-Instruct_bu_2to1/final/capability_lora.safetensors
dapo bu_1to2 configs/eval/eval_trainset_qwen25_dapo_lora_bu_1to2.yaml checkpoints/dapo/Qwen2.5-7B-Instruct_bu_1to2/final/capability_lora.safetensors
dapo 1to2    configs/eval/eval_trainset_qwen25_dapo_lora_1to2.yaml     checkpoints/dapo/Qwen2.5-7B-Instruct_1to2/final/capability_lora.safetensors
dapo bu_2to1 configs/eval/eval_trainset_qwen25_dapo_full_bu_2to1.yaml checkpoints/dapo_full/Qwen2.5-7B-Instruct_bu_2to1/final/done.flag
dapo bu_1to2 configs/eval/eval_trainset_qwen25_dapo_full_bu_1to2.yaml checkpoints/dapo_full/Qwen2.5-7B-Instruct_bu_1to2/final/done.flag
dapo 1to2    configs/eval/eval_trainset_qwen25_dapo_full_1to2.yaml     checkpoints/dapo_full/Qwen2.5-7B-Instruct_1to2/final/done.flag
EOF
)

mkdir -p debug_logs data/eval
n_run=0 n_skip=0
while read -r stage tag cfg ckpt; do
  [ -z "$stage" ] && continue
  [ -n "$ONLY" ] && [ "$stage" != "$ONLY" ] && continue
  [ -n "$ONLY_TAG" ] && [ "$tag" != "$ONLY_TAG" ] && continue
  if [ ! -e "$ckpt" ]; then
    echo "[eval] SKIP (checkpoint 未就绪): $cfg  ← 缺 $ckpt"
    n_skip=$((n_skip+1)); continue
  fi
  if [ ! -f "$cfg" ]; then
    echo "[eval] SKIP (config 缺失): $cfg"; n_skip=$((n_skip+1)); continue
  fi
  name=$(basename "$cfg" .yaml)
  log="debug_logs/${name}.log"
  echo "[eval] === RUN $name  (stage=$stage split=$tag) ==="
  echo "[eval]     ckpt OK: $ckpt"
  echo "[eval]     log → $log"
  "$PY" scripts/train/11_eval.py --config "$cfg" $JUDGE_FLAG 2>&1 | tee -a "$log"
  n_run=$((n_run+1))
done <<< "$ROWS"

echo "[eval] ===== DONE: ran $n_run, skipped $n_skip ====="
echo "[eval] 逐 checkpoint 结果在 data/eval/<name>.stats.json"
echo "[eval] 跑 seen/unseen 拆分 + 横向对比表:"
echo "       $PY scripts/train/analyze_eval_splits.py"