YUGOROU commited on
Commit
68f8501
·
verified ·
1 Parent(s): 94e32ba

sync: eqbench-ja-run/README.md

Browse files
Files changed (1) hide show
  1. eqbench-ja-run/README.md +88 -0
eqbench-ja-run/README.md ADDED
@@ -0,0 +1,88 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # EQ-Bench3 ローカル評価セットアップ
2
+
3
+ TeenEmo-LFM2.5-1.2B-DPO を EQ-Bench3 日本語版で評価するセットアップ。
4
+ 受験者・採点者ともに同一A100インスタンス上のローカルvLLMを使用。
5
+
6
+ ## モデル構成
7
+
8
+ | 役割 | モデル | VRAM | ポート |
9
+ |------|--------|------|--------|
10
+ | 受験者 | `YUGOROU/TeenEmo-LFM2.5-1.2B-DPO` | ~3GB | 8000 |
11
+ | 採点者 | `Qwen/Qwen3.5-35B-A3B` | ~70GB | 8001 |
12
+
13
+ ## セットアップ
14
+
15
+ ```bash
16
+ export HF_TOKEN="hf_xxxx" && export HF_USERNAME="YUGOROU"
17
+ curl -fL -H "Authorization: Bearer ${HF_TOKEN}" \
18
+ "https://huggingface.co/datasets/YUGOROU/Test-2/resolve/main/eqbench-ja-run/setup_eqbench_run.sh" \
19
+ -o /tmp/setup_eqbench_run.sh && bash /tmp/setup_eqbench_run.sh
20
+ ```
21
+
22
+ ## 同時起動モード(推奨)
23
+
24
+ A100 80GBで両モデルを同時起動する。TeenEmoのGPU_UTIL=0.10で約3GBに抑える。
25
+
26
+ ```bash
27
+ # Step 1: TeenEmo起動(port 8000)
28
+ tmux new-session -d -s eq_run
29
+ tmux new-window -t eq_run -n test
30
+ tmux send-keys -t eq_run:test "cd /workspace/eqbench-run && export HF_TOKEN='hf_xxxx' && ./serve_test.sh" Enter
31
+
32
+ # Step 2: 採点者起動(port 8001)
33
+ tmux new-window -t eq_run -n judge
34
+ tmux send-keys -t eq_run:judge "cd /workspace/eqbench-run && ./serve_judge.sh" Enter
35
+
36
+ # Step 3: 起動確認
37
+ tmux capture-pane -t eq_run:test -p | grep "startup complete"
38
+ tmux capture-pane -t eq_run:judge -p | grep "startup complete"
39
+
40
+ # Step 4: 評価実行
41
+ cd /workspace/eqbench-run/eqbench3
42
+ python eqbench3.py \
43
+ --test-model YUGOROU/TeenEmo-LFM2.5-1.2B-DPO \
44
+ --model-name TeenEmo-DPO \
45
+ --judge-model Qwen/Qwen3.5-35B-A3B \
46
+ --no-elo \
47
+ --iterations 1
48
+ ```
49
+
50
+ ## 順次実行モード(OOM対策)
51
+
52
+ 同時起動でOOMが発生した場合。
53
+
54
+ ```bash
55
+ # Phase 1: TeenEmoで全応答を生成
56
+ ./serve_test.sh & # port 8000
57
+ # (起動待機)
58
+ cd /workspace/eqbench-run/eqbench3
59
+ python eqbench3.py \
60
+ --test-model YUGOROU/TeenEmo-LFM2.5-1.2B-DPO \
61
+ --model-name TeenEmo-DPO \
62
+ --judge-model Qwen/Qwen3.5-35B-A3B \
63
+ --no-elo --iterations 1
64
+ # → 採点時にjudge APIが返せずエラーになるが応答データは保存される
65
+
66
+ # Phase 2: TeenEmoを停止してJudgeで採点
67
+ kill $(lsof -t -i:8000)
68
+ JUDGE_GPU_UTIL=0.90 ./serve_judge.sh & # port 8001
69
+ # (起動待機後、再実行で既存応答を再利用し採点のみ行う)
70
+ python eqbench3.py \
71
+ --test-model YUGOROU/TeenEmo-LFM2.5-1.2B-DPO \
72
+ --model-name TeenEmo-DPO \
73
+ --judge-model Qwen/Qwen3.5-35B-A3B \
74
+ --no-elo --iterations 1
75
+ ```
76
+
77
+ ## 結果の確認
78
+
79
+ ```bash
80
+ cat /workspace/eqbench-run/eqbench3/eqbench3_runs.json | python3 -c "
81
+ import json, sys
82
+ data = json.load(sys.stdin)
83
+ for run_id, run in data.items():
84
+ if 'TeenEmo' in run_id:
85
+ print(f'Run: {run_id}')
86
+ print(f'Score: {run.get(\"eq_bench_score\", \"N/A\")}')
87
+ "
88
+ ```