| # EQ-Bench3 日本語版 ローカル評価セットアップ |
|
|
| TeenEmo-LFM2.5-1.2B-DPO を EQ-Bench3 日本語版で評価するセットアップ。 |
| 受験者・採点者ともに同一インスタンス上のローカルvLLMを使用。 |
|
|
| ## 修正済み問題 |
|
|
| | # | 重大度 | 問題 | 修正 | |
| |---|--------|------|------| |
| | JA-D1 | 🔴 Fatal | scenario_master_prompt 等 9 ファイルが差し替えられていなかった(scenario_prompts / notes の 2 ファイルのみ) | 全 12 ファイルを差し替え | |
| | JA-D2 | 🔴 Fatal | benchmark.py / conversation.py のコードパッチ未適用 → スコア全0 | setup 時に patch_benchmark.py / patch_conversation.py を自動適用 | |
| | JA-M1 | 🔴 Fatal | serve_test.sh の MAX_MODEL_LEN=4096 → EQ-Bench3 は max_tokens=12000 を要求するためクラッシュ | 32768 に修正 | |
| |
| ## モデル構成 |
| |
| | 役割 | モデル | VRAM | ポート | |
| |------|--------|------|--------| |
| | 受験者 | `LiquidAI/LFM2.5-1.2B-Base` + LoRA `YUGOROU/TeenEmo-LFM2.5-1.2B-DPO` | ~3GB (GPU_UTIL=0.10) | 8000 | |
| | 採点者 | `Qwen/Qwen3.5-35B-A3B` | ~70GB (GPU_UTIL=0.88) | 8001 | |
| |
| ## セットアップ |
| |
| ```bash |
| export HF_TOKEN="hf_xxxx" |
| curl -fL -H "Authorization: Bearer ${HF_TOKEN}" \ |
| "https://huggingface.co/datasets/YUGOROU/Test-2/resolve/main/eqbench-ja-run/setup_eqbench_run.sh" \ |
| -o /tmp/setup_eqbench_run.sh && bash /tmp/setup_eqbench_run.sh |
| ``` |
| |
| ## 同時起動モード(推奨: A100 80GB) |
| |
| ```bash |
| # Step 1: TeenEmo 起動(port 8000, GPU_UTIL=0.10) |
| tmux new-session -d -s eq_run |
| tmux new-window -t eq_run -n test |
| tmux send-keys -t eq_run:test "cd /workspace/eqbench-run && export HF_TOKEN='$HF_TOKEN' && ./serve_test.sh" Enter |
| |
| # Step 2: Judge 起動(port 8001, GPU_UTIL=0.88) |
| tmux new-window -t eq_run -n judge |
| tmux send-keys -t eq_run:judge "cd /workspace/eqbench-run && ./serve_judge.sh" Enter |
| |
| # Step 3: 起動確認 |
| tmux capture-pane -t eq_run:test -p | grep "startup complete" |
| tmux capture-pane -t eq_run:judge -p | grep "startup complete" |
| |
| # Step 4: 評価実行 |
| cd /workspace/eqbench-run/eqbench3 |
| python eqbench3.py \ |
| --test-model teenemo-dpo \ |
| --model-name TeenEmo-DPO \ |
| --judge-model Qwen/Qwen3.5-35B-A3B \ |
| --no-elo \ |
| --save-interval 1 \ |
| --iterations 1 |
| ``` |
| |
| ## 順次実行モード(OOM 対策) |
| |
| `--save-interval 1` で 1 タスクごとに保存。サーバー切り替え後に再実行すると完了済みタスクをスキップして再開できる。 |
| |
| ```bash |
| # Phase 1: TeenEmo で応答生成(port 8000 のみ / Judge は失敗でOK) |
| TEST_GPU_UTIL=0.95 ./serve_test.sh & |
| cd /workspace/eqbench-run/eqbench3 |
| python eqbench3.py \ |
| --test-model teenemo-dpo --model-name TeenEmo-DPO \ |
| --judge-model Qwen/Qwen3.5-35B-A3B \ |
| --no-elo --save-interval 1 --iterations 1 |
| # Judge API 失敗は想定内。応答済みタスクは保存される |
|
|
| # Phase 2: TeenEmo 停止 → Judge 起動(port 8001)で採点のみ |
| pkill -f "vllm serve LiquidAI" || kill $(lsof -t -i:8000) || true |
| JUDGE_GPU_UTIL=0.90 ./serve_judge.sh & |
| python eqbench3.py \ |
| --test-model teenemo-dpo --model-name TeenEmo-DPO \ |
| --judge-model Qwen/Qwen3.5-35B-A3B \ |
| --no-elo --save-interval 1 --iterations 1 |
| ``` |
| |
| ## 結果確認 |
| |
| ```bash |
| cat /workspace/eqbench-run/eqbench3/eqbench3_runs.json | python3 -c " |
| import json, sys |
| data = json.load(sys.stdin) |
| for run_id, run in data.items(): |
| if 'TeenEmo' in run_id: |
| print('Run:', run_id) |
| print('Score:', run.get('eq_bench_score', 'N/A')) |
| " |
| ``` |
| |