# EQ-Bench3 日本語版 Vast.ai セットアップ TeenEmo-LFM2.5-1.2B-DPO を EQ-Bench3 日本語版で評価する Vast.ai セットアップ。 採点者は HF Inference Providers (novita / gpt-oss-120b) を使用。 ## 修正済み問題 | # | 重大度 | 問題 | 修正 | |---|--------|------|------| | JA-D1 | 🔴 Fatal | scenario_master_prompt 等 10 ファイルが差し替えられていなかった | 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 | patch_benchmark.py / patch_conversation.py を自動適用 | コードパッチの詳細は [teememo-eq-bench-ja](https://huggingface.co/datasets/YUGOROU/teememo-eq-bench-ja) の `patch/` を参照。 ## セットアップ ```bash export HF_TOKEN="hf_xxxx" curl -fL -H "Authorization: Bearer ${HF_TOKEN}" \ "https://huggingface.co/datasets/YUGOROU/Test-2/resolve/main/eqbench-vast/setup_eqbench_vast.sh" \ -o /tmp/setup_eqbench_vast.sh && bash /tmp/setup_eqbench_vast.sh ``` ## 実行 ```bash # Step 1: TeenEmo 起動 tmux new-session -d -s eq_run tmux send-keys -t eq_run "cd /workspace/eqbench-run && export HF_TOKEN='$HF_TOKEN' && ./serve_test_vast.sh" Enter # Step 2: 起動確認(Application startup complete. まで待つ) tmux capture-pane -t eq_run -p | tail -5 # Step 3: 評価実行 cd /workspace/eqbench-run/eqbench3 python eqbench3.py \ --test-model teenemo-dpo \ --model-name TeenEmo-DPO \ --judge-model openai/gpt-oss-120b \ --no-elo \ --save-interval 1 \ --iterations 1 ```