eqbench-run-v2/setup_eqbench_vast.sh

#!/usr/bin/env bash
# =============================================================================
# setup_eqbench_vast.sh — EQ-Bench3 日本語版セットアップ (Vast.ai)
#
# 構成:
#   受験者: TeenEmo-LFM2.5-1.2B-DPO (LoRA) on Vast.ai vLLM (port 8000)
#   採点者: openai/gpt-oss-120b via HF Inference Providers / novita
#
# 修正履歴:
#   [Fix-1] max_lora_rank=32: Unsloth LoRA は rank=32
#   [Fix-2] max_model_len=32768: EQ-Bench3 の max_tokens=12000 に対応
#   [Fix-3] HF Inference Providers の Judge URL: router.huggingface.co/novita
#   [Fix-4] --save-interval 1: 途中再開を保証
#   [Fix-5] MAX_RETRIES=6, REQUEST_TIMEOUT=300: HF API レート制限・遅延対応
#   [Fix-JA-D1] 全12日本語データファイルを差し替え（一部だけでは不十分）
#   [Fix-JA-D2] benchmark.py / conversation.py にコードパッチを適用
# =============================================================================

set -euo pipefail

HF_TOKEN="${HF_TOKEN:?HF_TOKEN が未設定です。export HF_TOKEN='hf_xxxx'}"
HF_USERNAME="${HF_USERNAME:-YUGOROU}"
WORKSPACE="/workspace/eqbench-run"
TEST_PORT="${TEST_PORT:-8000}"
LORA_NAME="${TEST_LORA_NAME:-teenemo-dpo}"

PATCH_REPO="${PATCH_REPO:-${HF_USERNAME}/teememo-eq-bench-ja}"
SCRIPTS_REPO="${SCRIPTS_REPO:-${HF_USERNAME}/Test-2}"

echo "[setup] EQ-Bench3 日本語版 Vast.ai セットアップ"
echo "[setup] 採点者: HF Inference Providers / novita / gpt-oss-120b"
echo ""

# ── 依存パッケージ ────────────────────────────────────────────
echo "[setup] 依存パッケージのインストール..."
pip install -q requests python-dotenv tqdm numpy scipy trueskill 2>/dev/null || true
apt-get install -y tmux git 2>/dev/null || true

# ── EQ-Bench3 クローン ────────────────────────────────────────
mkdir -p "${WORKSPACE}"
if [ -d "${WORKSPACE}/eqbench3" ]; then
    echo "[setup] EQ-Bench3 更新中..."
    git -C "${WORKSPACE}/eqbench3" pull --ff-only 2>/dev/null || true
else
    echo "[setup] EQ-Bench3 クローン中..."
    git clone --depth=1 https://github.com/EQ-bench/eqbench3.git "${WORKSPACE}/eqbench3"
fi
echo "[setup] ✅ EQ-Bench3"

# ── [Fix-JA-D1] 全12日本語データファイルを差し替え ──────────
echo ""
echo "[setup] [Fix-JA-D1] 全12日本語データファイルをDL中..."
JA_COUNT=0
for pair in \
    "scenario_prompts_ja.txt|scenario_prompts.txt" \
    "scenario_notes_ja.txt|scenario_notes.txt" \
    "scenario_master_prompt_ja.txt|scenario_master_prompt.txt" \
    "scenario_master_prompt_message_drafting_ja.txt|scenario_master_prompt_message_drafting.txt" \
    "scenario_master_prompt_analysis_ja.txt|scenario_master_prompt_analysis.txt" \
    "rubric_scoring_prompt_ja.txt|rubric_scoring_prompt.txt" \
    "rubric_scoring_prompt_analysis_ja.txt|rubric_scoring_prompt_analysis.txt" \
    "rubric_scoring_criteria_ja.txt|rubric_scoring_criteria.txt" \
    "rubric_scoring_criteria_analysis_ja.txt|rubric_scoring_criteria_analysis.txt" \
    "pairwise_prompt_eqbench3_ja.txt|pairwise_prompt_eqbench3.txt" \
    "pairwise_prompt_eqbench3_analysis_ja.txt|pairwise_prompt_eqbench3_analysis.txt" \
    "debrief_prompt_ja.txt|debrief_prompt.txt"; do
    SRC="${pair%%|*}"
    DEST="${WORKSPACE}/eqbench3/data/${pair##*|}"
    [ -f "${DEST}" ] && [ ! -f "${DEST}.en.bak" ] && cp "${DEST}" "${DEST}.en.bak"
    if curl -sfL \
        -H "Authorization: Bearer ${HF_TOKEN}" \
        "https://huggingface.co/datasets/${PATCH_REPO}/resolve/main/data/${SRC}" \
        -o "${DEST}"; then
        echo "[setup]   ✅ ${pair##*|}"
        JA_COUNT=$((JA_COUNT + 1))
    else
        [ -f "${DEST}.en.bak" ] && cp "${DEST}.en.bak" "${DEST}"
        echo "[setup]   ⚠️  DL失敗 → 英語版: ${pair##*|}"
    fi
done
echo "[setup] 日本語データファイル: ${JA_COUNT}/12"

# ── [Fix-JA-D2] コードパッチ (benchmark.py / conversation.py) ─
echo ""
echo "[setup] [Fix-JA-D2] コードパッチを適用中..."
PATCH_DIR="${WORKSPACE}/eqbench-patch"
mkdir -p "${PATCH_DIR}"
PATCH_OK=0
for pf in patch_benchmark.py patch_conversation.py; do
    if curl -sfL \
        -H "Authorization: Bearer ${HF_TOKEN}" \
        "https://huggingface.co/datasets/${PATCH_REPO}/resolve/main/patch/${pf}" \
        -o "${PATCH_DIR}/${pf}"; then
        PATCH_OK=$((PATCH_OK + 1))
    else
        echo "[setup]   ⚠️  DL失敗: ${pf}"
    fi
done

if [ "${PATCH_OK}" -eq 2 ]; then
    cd "${WORKSPACE}/eqbench3"
    python3 "${PATCH_DIR}/patch_benchmark.py"
    python3 "${PATCH_DIR}/patch_conversation.py"
    echo "[setup]   ✅ コードパッチ完了"
else
    echo "[setup]   ❌ パッチファイル取得失敗。PATCH_REPO=${PATCH_REPO} を確認してください。"
fi

# ── serve_test_vast.sh DL ────────────────────────────────────
curl -fL \
    -H "Authorization: Bearer ${HF_TOKEN}" \
    "https://huggingface.co/datasets/${SCRIPTS_REPO}/resolve/main/eqbench-vast/serve_test_vast.sh" \
    -o "${WORKSPACE}/serve_test_vast.sh"
chmod +x "${WORKSPACE}/serve_test_vast.sh"
echo "[setup] ✅ serve_test_vast.sh"

# ── .env 生成 ────────────────────────────────────────────────
cat > "${WORKSPACE}/eqbench3/.env" << ENVEOF
# EQ-Bench3 日本語版設定 (Vast.ai + HF Inference Providers)
TEST_API_URL=http://localhost:${TEST_PORT}/v1/chat/completions
TEST_API_KEY=dummy
JUDGE_API_URL=https://router.huggingface.co/novita/v1/chat/completions
JUDGE_API_KEY=${HF_TOKEN}
MAX_RETRIES=6
RETRY_DELAY=10
REQUEST_TIMEOUT=300
ENVEOF
echo "[setup] ✅ .env 生成完了"

echo ""
echo "[setup] =============================="
echo "[setup] ✅ セットアップ完了"
echo "[setup] =============================="
echo ""
echo "# Step 1: TeenEmo サーバー起動"
echo "tmux new-session -d -s eq_run"
echo "tmux send-keys -t eq_run \"cd ${WORKSPACE} && export HF_TOKEN='${HF_TOKEN}' && ./serve_test_vast.sh\" Enter"
echo ""
echo "# Step 2: 起動確認"
echo "tmux capture-pane -t eq_run -p | tail -5"
echo ""
echo "# Step 3: EQ-Bench3 実行"
echo "cd ${WORKSPACE}/eqbench3"
echo "python eqbench3.py --test-model ${LORA_NAME} --model-name TeenEmo-DPO --judge-model openai/gpt-oss-120b --no-elo --save-interval 1 --iterations 1"