| # HumanEval+ / MBPP+ 评测 (EvalPlus, vLLM 后端) | |
| # 官方项目: https://github.com/evalplus/evalplus | |
| # 用法: bash run_evalplus.sh [MODEL_PATH] [TP] | |
| # TP 不传或传 auto 时自动使用当前可见的全部 GPU 数 | |
| # 结果输出到: /data/icto/mc56499/code_eval/evalplus_results/ | |
| # samples: evalplus_results/<dataset>/<model>_vllm_temp_0.0.jsonl | |
| # 报告: 同目录 *_eval_results.json (含 pass@1 base/plus) | |
| set -e | |
| MODEL_PATH=${1:-/data/icto/mc56499/model/Qwen2.5-Coder-3B} | |
| TP=${2:-auto} | |
| if [ "$TP" = "auto" ] || [ -z "$TP" ]; then | |
| TP=$(nvidia-smi -L | wc -l) | |
| fi | |
| ROOT=/data/icto/mc56499/code_eval/evalplus_results | |
| PARALLEL=16 # 评测进程数 | |
| mkdir -p "$ROOT" | |
| for DATASET in humaneval mbpp; do | |
| echo "===== [EvalPlus] dataset=$DATASET model=$MODEL_PATH tp=$TP =====" | |
| evalplus.evaluate \ | |
| --model "$MODEL_PATH" \ | |
| --dataset "$DATASET" \ | |
| --backend vllm \ | |
| --tp "$TP" \ | |
| --greedy \ | |
| --force_base_prompt \ | |
| --root "$ROOT" \ | |
| --parallel "$PARALLEL" | |
| done | |
| echo "===== EvalPlus 结果汇总 =====" | |
| for f in $(find "$ROOT" -name "*_eval_results.json" | sort); do | |
| echo "--- $f" | |
| python - "$f" <<'EOF' | |
| import json, sys | |
| with open(sys.argv[1]) as f: | |
| r = json.load(f) | |
| tasks = r["eval"].values() | |
| base = sum(1 for samples in tasks if any(s["base_status"] == "pass" for s in samples)) | |
| plus = sum(1 for samples in r["eval"].values() if any( | |
| s["base_status"] == "pass" and s["plus_status"] == "pass" for s in samples)) | |
| n = len(r["eval"]) | |
| print(f" pass@1 (base tests): {base / n:.4f} ({base}/{n})") | |
| print(f" pass@1 (base+extra): {plus / n:.4f} ({plus}/{n})") | |
| EOF | |
| done | |