Create standalone static benchmark leaderboard

.gitignore

@@ -0,0 +1,6 @@
+__pycache__/
+*.py[cod]
+.pytest_cache/
+.env
+.env.*
+.venv/

README.md

@@ -0,0 +1,93 @@
+---
+title: KETI Benchmark Leaderboard
+emoji: 🦀
+colorFrom: red
+colorTo: gray
+sdk: static
+app_file: index.html
+pinned: false
+---
+
+# KETI Leaderboard
+
+A static leaderboard for published KETI model evaluations. The browser loads
+`front/data/benchmark.json` directly; no Python server, Docker runtime, access token,
+or build step is needed. Dataset selection, model search, provider filtering,
+comparison charts, theme selection, and CSV export are available.
+
+The initial snapshot contains the 19 models and the 9 displayed datasets retrieved
+from the existing public leaderboard during migration. Scores and evaluation
+timestamps are preserved. The snapshot records its source and capture time;
+internal evaluation job identifiers and error messages are not included.
+
+## Run locally
+
+From the repository root:
+
+```bash
+python3 -m http.server 7860 --bind 127.0.0.1
+```
+
+Open <http://127.0.0.1:7860/>. Use an HTTP server rather than opening `index.html`
+as a local file, because the browser loads JavaScript modules and JSON.
+Tailwind, Chart.js, and fonts currently load from their existing public CDNs.
+
+## Verify the static page
+
+With Playwright and Chrome/Chromium installed, run:
+
+```bash
+python3 -m unittest discover -s tests -v
+```
+
+Set `CHROME_PATH` if the browser is not on `PATH`. These checks use a temporary
+local HTTP server and cover published scores and means, filters, comparison charts,
+CSV export, mobile controls, preferences, loading failures, and malformed data.
+Internet access is needed for the existing CDN assets in the chart test.
+
+## Update published results
+
+1. Evaluate models outside this Space.
+2. Update `front/data/benchmark.json` with the resulting scores.
+3. Commit and push the changed data file to the Space. Visitors can use **Refresh**
+   to load the published version again.
+
+The data file accepts either the legacy `benchmark.json` array of models or an
+object with `models` and an optional `datasets` array. Each model has `provider`,
+`name`, `repo`, `is_multimodal`, `updated_at`, and `scores`. Each score has
+`dataset_name`, numeric `score`, and `metric_type` (`raw` or `llm-as-judge`).
+The `datasets` array sets the displayed columns and their order; without it,
+the canonical datasets and any additional scored datasets are shown.
+
+The mean calculation retains the previous behavior: all selected, supported
+datasets must have scores. Unsupported multimodal datasets are excluded for
+text-only models. Missing means are displayed as `—`. A failed data request shows
+an error; it never substitutes generated scores. A failed refresh keeps the last
+successfully loaded results on screen.
+
+For an independently updated, **public** JSON source, change `dataUrl` in
+`front/config.mjs` to its HTTPS URL. That server must allow browser cross-origin
+requests (CORS). No automatic fallback to another source is applied. The default
+bundled snapshot remains independent of the personal Hugging Face account.
+All browser configuration and bundled files are public: do not put tokens or
+private evaluation data in them.
+
+## Deploy as a new Hugging Face Space
+
+This directory is an independent Git repository containing only the static app.
+The existing Docker Space at
+https://huggingface.co/spaces/DoolyKim22/KETI_Leaderboard remains separate.
+
+The deployment target is the new organization-owned Space
+[KETI-NLP/Benchmark_Leaderboard](https://huggingface.co/spaces/KETI-NLP/Benchmark_Leaderboard),
+using the **Static** SDK. The README metadata already specifies
+`sdk: static` and `app_file: index.html` as described in the
+[Static Spaces documentation](https://huggingface.co/docs/hub/spaces-sdks-static).
+Do not use the existing Docker Space as this repository's remote.
+
+The snapshot is independent of the running Docker app. Updates published to the
+old app after the snapshot was captured are not automatically copied here.
+Update `front/data/benchmark.json` and publish a new commit to refresh this app.
+
+Model submissions, automatic evaluations, and dataset backup workers are not part
+of the static site. Browser configuration lives in `front/config.mjs`.

front/config.mjs

@@ -0,0 +1,6 @@
+// Public browser configuration. Never put Hugging Face tokens here.
+export const config = {
+  // Relative to this file. An absolute, public HTTPS JSON URL also works.
+  dataUrl: new URL('./data/benchmark.json', import.meta.url).href,
+  requestTimeoutMs: 15000,
+};

front/data.mjs

@@ -0,0 +1,81 @@
+export const CANONICAL_DATASETS = [
+  '일반 상식', '멀티 모달', '수치 정보', '팩트 체크', '소버린 벤치마크',
+  'K-Prism (Text, Intact)', 'K-Prism (Text, Perturbed)',
+  'K-Prism (Image, Intact)', 'K-Prism (Image, Perturbed)',
+];
+
+const isText = value => typeof value === 'string' && value.trim().length > 0;
+
+// Accept both the original benchmark.json array and a published {models} snapshot.
+// Validate before replacing the visible results so failed refreshes keep good data.
+export function normalizeBenchmark(payload) {
+  const source = Array.isArray(payload) ? payload : payload?.models;
+  if (!Array.isArray(source)) throw new Error('Expected a model array or an object with models.');
+
+  const modelKeys = new Set();
+  const models = source.map((model, index) => {
+    if (!model || !isText(model.provider) || !isText(model.name)) {
+      throw new Error(`Model ${index + 1} is missing its provider or name.`);
+    }
+    const key = JSON.stringify([model.provider, model.name]);
+    if (modelKeys.has(key)) throw new Error(`Duplicate model: ${model.provider} / ${model.name}.`);
+    modelKeys.add(key);
+    if (!Array.isArray(model.scores)) throw new Error(`Invalid scores for ${model.name}.`);
+    if (model.is_multimodal != null && typeof model.is_multimodal !== 'boolean') {
+      throw new Error(`Invalid model type for ${model.name}.`);
+    }
+    const scoreNames = new Set();
+    const scores = model.scores.map(score => {
+      if (!score || !isText(score.dataset_name) || !Number.isFinite(score.score)) {
+        throw new Error(`Invalid score for ${model.name}.`);
+      }
+      if (scoreNames.has(score.dataset_name)) {
+        throw new Error(`Duplicate dataset score for ${model.name}: ${score.dataset_name}.`);
+      }
+      scoreNames.add(score.dataset_name);
+      return {
+        dataset_name: score.dataset_name,
+        metric_type: score.metric_type === 'llm-as-judge' ? 'llm-as-judge' : 'raw',
+        score: score.score,
+      };
+    });
+    return {
+      provider: model.provider,
+      name: model.name,
+      repo: typeof model.repo === 'string' ? model.repo : '',
+      is_multimodal: model.is_multimodal ?? false,
+      status: typeof model.status === 'string' ? model.status : 'completed',
+      updated_at: typeof model.updated_at === 'string' ? model.updated_at : '',
+      scores,
+    };
+  });
+
+  const publishedDatasets = Array.isArray(payload) ? undefined : payload.datasets;
+  if (publishedDatasets !== undefined &&
+      (!Array.isArray(publishedDatasets) || !publishedDatasets.every(isText))) {
+    throw new Error('Invalid dataset list.');
+  }
+  // A published list preserves the curator's selection and order. Legacy arrays
+  // include the canonical datasets plus any additional scored datasets.
+  const datasets = [...new Set(publishedDatasets ?? [
+    ...CANONICAL_DATASETS,
+    ...models.flatMap(model => model.scores.map(score => score.dataset_name)),
+  ])];
+  return { models, datasets };
+}
+
+export async function loadBenchmark(url, timeoutMs = 15000) {
+  const controller = new AbortController();
+  const timeout = setTimeout(() => controller.abort(), timeoutMs);
+  try {
+    const response = await fetch(url, {
+      cache: 'no-store',
+      credentials: 'omit',
+      signal: controller.signal,
+    });
+    if (!response.ok) throw new Error(`Score data request failed (HTTP ${response.status}).`);
+    return normalizeBenchmark(await response.json());
+  } finally {
+    clearTimeout(timeout);
+  }
+}

front/data/benchmark.json

@@ -0,0 +1,962 @@
+{
+  "source": "https://doolykim22-keti-leaderboard.hf.space/api/results",
+  "captured_at": "2026-09-17T06:05:24+00:00",
+  "datasets": [
+    "일반 상식",
+    "멀티 모달",
+    "수치 정보",
+    "팩트 체크",
+    "소버린 벤치마크",
+    "K-Prism (Text, Intact)",
+    "K-Prism (Text, Perturbed)",
+    "K-Prism (Image, Intact)",
+    "K-Prism (Image, Perturbed)"
+  ],
+  "models": [
+    {
+      "provider": "Alibaba Cloud",
+      "name": "qwen3-30b",
+      "repo": "Qwen/Qwen3-30B-A3B-Thinking-2507",
+      "is_multimodal": false,
+      "status": "completed",
+      "updated_at": "2026-06-26T06:07:08Z",
+      "scores": [
+        {
+          "dataset_name": "일반 상식",
+          "metric_type": "raw",
+          "score": 49.22
+        },
+        {
+          "dataset_name": "수치 정보",
+          "metric_type": "raw",
+          "score": 43.33
+        },
+        {
+          "dataset_name": "특수 도메인",
+          "metric_type": "raw",
+          "score": 100.0
+        },
+        {
+          "dataset_name": "소버린 벤치마크",
+          "metric_type": "raw",
+          "score": 62.14
+        }
+      ]
+    },
+    {
+      "provider": "Alibaba Cloud",
+      "name": "qwen3-32b",
+      "repo": "Qwen/Qwen3-32B",
+      "is_multimodal": false,
+      "status": "completed",
+      "updated_at": "2026-06-26T07:06:54Z",
+      "scores": [
+        {
+          "dataset_name": "일반 상식",
+          "metric_type": "raw",
+          "score": 38.86
+        },
+        {
+          "dataset_name": "수치 정보",
+          "metric_type": "raw",
+          "score": 42.38
+        },
+        {
+          "dataset_name": "특수 도메인",
+          "metric_type": "raw",
+          "score": 100.0
+        },
+        {
+          "dataset_name": "소버린 벤치마크",
+          "metric_type": "raw",
+          "score": 58.64
+        }
+      ]
+    },
+    {
+      "provider": "Alibaba Cloud",
+      "name": "qwen3-4b",
+      "repo": "Qwen/Qwen3-4B-Thinking-2507",
+      "is_multimodal": false,
+      "status": "completed",
+      "updated_at": "2026-06-26T07:22:21Z",
+      "scores": [
+        {
+          "dataset_name": "일반 상식",
+          "metric_type": "raw",
+          "score": 38.86
+        },
+        {
+          "dataset_name": "수치 정보",
+          "metric_type": "raw",
+          "score": 34.76
+        },
+        {
+          "dataset_name": "특수 도메인",
+          "metric_type": "raw",
+          "score": 100.0
+        },
+        {
+          "dataset_name": "소버린 벤치마크",
+          "metric_type": "raw",
+          "score": 36.32
+        }
+      ]
+    },
+    {
+      "provider": "Alibaba Cloud",
+      "name": "qwen3-8b",
+      "repo": "Qwen/Qwen3-8B",
+      "is_multimodal": false,
+      "status": "completed",
+      "updated_at": "2026-06-26T07:38:00Z",
+      "scores": [
+        {
+          "dataset_name": "일반 상식",
+          "metric_type": "raw",
+          "score": 10.36
+        },
+        {
+          "dataset_name": "수치 정보",
+          "metric_type": "raw",
+          "score": 33.81
+        },
+        {
+          "dataset_name": "특수 도메인",
+          "metric_type": "raw",
+          "score": 50.0
+        },
+        {
+          "dataset_name": "소버린 벤치마크",
+          "metric_type": "raw",
+          "score": 26.91
+        }
+      ]
+    },
+    {
+      "provider": "Alibaba Cloud",
+      "name": "qwen3-vl-4b",
+      "repo": "Qwen/Qwen3-VL-4B-Thinking",
+      "is_multimodal": true,
+      "status": "completed",
+      "updated_at": "2026-06-26T07:49:35Z",
+      "scores": [
+        {
+          "dataset_name": "일반 상식",
+          "metric_type": "raw",
+          "score": 25.91
+        },
+        {
+          "dataset_name": "멀티 모달",
+          "metric_type": "raw",
+          "score": 28.21
+        },
+        {
+          "dataset_name": "수치 정보",
+          "metric_type": "raw",
+          "score": 19.52
+        },
+        {
+          "dataset_name": "팩트 체크",
+          "metric_type": "raw",
+          "score": 68.57
+        },
+        {
+          "dataset_name": "K-Prism (Text, Intact)",
+          "metric_type": "raw",
+          "score": 60.32
+        },
+        {
+          "dataset_name": "K-Prism (Text, Perturbed)",
+          "metric_type": "raw",
+          "score": 54.37
+        },
+        {
+          "dataset_name": "K-Prism (Image, Intact)",
+          "metric_type": "raw",
+          "score": 89.16
+        },
+        {
+          "dataset_name": "K-Prism (Image, Perturbed)",
+          "metric_type": "raw",
+          "score": 23.49
+        },
+        {
+          "dataset_name": "소버린 벤치마크",
+          "metric_type": "raw",
+          "score": 20.35
+        }
+      ]
+    },
+    {
+      "provider": "Alibaba Cloud",
+      "name": "qwen3-vl-8b",
+      "repo": "Qwen/Qwen3-VL-8B-Instruct",
+      "is_multimodal": true,
+      "status": "completed",
+      "updated_at": "2026-06-26T08:15:48Z",
+      "scores": [
+        {
+          "dataset_name": "일반 상식",
+          "metric_type": "raw",
+          "score": 1.55
+        },
+        {
+          "dataset_name": "멀티 모달",
+          "metric_type": "raw",
+          "score": 41.07
+        },
+        {
+          "dataset_name": "수치 정보",
+          "metric_type": "raw",
+          "score": 37.62
+        },
+        {
+          "dataset_name": "팩트 체크",
+          "metric_type": "raw",
+          "score": 71.43
+        },
+        {
+          "dataset_name": "K-Prism (Text, Intact)",
+          "metric_type": "raw",
+          "score": 61.31
+        },
+        {
+          "dataset_name": "K-Prism (Text, Perturbed)",
+          "metric_type": "raw",
+          "score": 61.31
+        },
+        {
+          "dataset_name": "K-Prism (Image, Intact)",
+          "metric_type": "raw",
+          "score": 80.32
+        },
+        {
+          "dataset_name": "K-Prism (Image, Perturbed)",
+          "metric_type": "raw",
+          "score": 21.69
+        },
+        {
+          "dataset_name": "소버린 벤치마크",
+          "metric_type": "raw",
+          "score": 3.94
+        }
+      ]
+    },
+    {
+      "provider": "Alibaba Cloud",
+      "name": "qwen3-vl-30b",
+      "repo": "Qwen/Qwen3-VL-30B-A3B-Thinking",
+      "is_multimodal": true,
+      "status": "completed",
+      "updated_at": "2026-06-29T00:55:39Z",
+      "scores": [
+        {
+          "dataset_name": "일반 상식",
+          "metric_type": "raw",
+          "score": 34.72
+        },
+        {
+          "dataset_name": "멀티 모달",
+          "metric_type": "raw",
+          "score": 42.14
+        },
+        {
+          "dataset_name": "수치 정보",
+          "metric_type": "raw",
+          "score": 46.19
+        },
+        {
+          "dataset_name": "팩트 체크",
+          "metric_type": "raw",
+          "score": 78.57
+        },
+        {
+          "dataset_name": "K-Prism (Text, Intact)",
+          "metric_type": "raw",
+          "score": 62.9
+        },
+        {
+          "dataset_name": "K-Prism (Text, Perturbed)",
+          "metric_type": "raw",
+          "score": 58.33
+        },
+        {
+          "dataset_name": "K-Prism (Image, Intact)",
+          "metric_type": "raw",
+          "score": 82.33
+        },
+        {
+          "dataset_name": "K-Prism (Image, Perturbed)",
+          "metric_type": "raw",
+          "score": 28.92
+        },
+        {
+          "dataset_name": "소버린 벤치마크",
+          "metric_type": "raw",
+          "score": 49.89
+        }
+      ]
+    },
+    {
+      "provider": "Alibaba Cloud",
+      "name": "qwen3-vl-32b",
+      "repo": "Qwen/Qwen3-VL-32B-Thinking",
+      "is_multimodal": true,
+      "status": "completed",
+      "updated_at": "2026-06-29T02:00:53Z",
+      "scores": [
+        {
+          "dataset_name": "일반 상식",
+          "metric_type": "raw",
+          "score": 23.83
+        },
+        {
+          "dataset_name": "멀티 모달",
+          "metric_type": "raw",
+          "score": 45.36
+        },
+        {
+          "dataset_name": "수치 정보",
+          "metric_type": "raw",
+          "score": 44.29
+        },
+        {
+          "dataset_name": "팩트 체크",
+          "metric_type": "raw",
+          "score": 82.86
+        },
+        {
+          "dataset_name": "K-Prism (Text, Intact)",
+          "metric_type": "raw",
+          "score": 63.89
+        },
+        {
+          "dataset_name": "K-Prism (Text, Perturbed)",
+          "metric_type": "raw",
+          "score": 60.32
+        },
+        {
+          "dataset_name": "K-Prism (Image, Intact)",
+          "metric_type": "raw",
+          "score": 69.68
+        },
+        {
+          "dataset_name": "K-Prism (Image, Perturbed)",
+          "metric_type": "raw",
+          "score": 28.11
+        },
+        {
+          "dataset_name": "소버린 벤치마크",
+          "metric_type": "raw",
+          "score": 22.54
+        }
+      ]
+    },
+    {
+      "provider": "Kakao",
+      "name": "kanana1.5",
+      "repo": "kakaocorp/kanana-1.5-v-3b-instruct",
+      "is_multimodal": true,
+      "status": "completed",
+      "updated_at": "2026-06-29T02:07:21Z",
+      "scores": [
+        {
+          "dataset_name": "일반 상식",
+          "metric_type": "raw",
+          "score": 49.22
+        },
+        {
+          "dataset_name": "멀티 모달",
+          "metric_type": "raw",
+          "score": 23.93
+        },
+        {
+          "dataset_name": "수치 정보",
+          "metric_type": "raw",
+          "score": 35.24
+        },
+        {
+          "dataset_name": "팩트 체크",
+          "metric_type": "raw",
+          "score": 58.57
+        },
+        {
+          "dataset_name": "특수 도메인",
+          "metric_type": "raw",
+          "score": 100.0
+        },
+        {
+          "dataset_name": "K-Prism (Text, Intact)",
+          "metric_type": "raw",
+          "score": 8.33
+        },
+        {
+          "dataset_name": "K-Prism (Text, Perturbed)",
+          "metric_type": "raw",
+          "score": 59.52
+        },
+        {
+          "dataset_name": "K-Prism (Image, Intact)",
+          "metric_type": "raw",
+          "score": 85.34
+        },
+        {
+          "dataset_name": "K-Prism (Image, Perturbed)",
+          "metric_type": "raw",
+          "score": 12.85
+        },
+        {
+          "dataset_name": "소버린 벤치마크",
+          "metric_type": "raw",
+          "score": 39.82
+        }
+      ]
+    },
+    {
+      "provider": "Shanghai AI Laboratory (OpenGVLab)",
+      "name": "InternVL3-8B",
+      "repo": "OpenGVLab/InternVL3-8B",
+      "is_multimodal": true,
+      "status": "completed",
+      "updated_at": "2026-06-29T02:19:52Z",
+      "scores": [
+        {
+          "dataset_name": "일반 상식",
+          "metric_type": "raw",
+          "score": 43.01
+        },
+        {
+          "dataset_name": "멀티 모달",
+          "metric_type": "raw",
+          "score": 22.86
+        },
+        {
+          "dataset_name": "수치 정보",
+          "metric_type": "raw",
+          "score": 43.81
+        },
+        {
+          "dataset_name": "팩트 체크",
+          "metric_type": "raw",
+          "score": 64.29
+        },
+        {
+          "dataset_name": "특수 도메인",
+          "metric_type": "raw",
+          "score": 100.0
+        },
+        {
+          "dataset_name": "K-Prism (Text, Intact)",
+          "metric_type": "raw",
+          "score": 60.52
+        },
+        {
+          "dataset_name": "K-Prism (Text, Perturbed)",
+          "metric_type": "raw",
+          "score": 59.72
+        },
+        {
+          "dataset_name": "K-Prism (Image, Intact)",
+          "metric_type": "raw",
+          "score": 66.67
+        },
+        {
+          "dataset_name": "K-Prism (Image, Perturbed)",
+          "metric_type": "raw",
+          "score": 39.76
+        },
+        {
+          "dataset_name": "소버린 벤치마크",
+          "metric_type": "raw",
+          "score": 42.67
+        }
+      ]
+    },
+    {
+      "provider": "NCSOFT",
+      "name": "varco",
+      "repo": "NCSOFT/VARCO-VISION-2.0-1.7B",
+      "is_multimodal": true,
+      "status": "completed",
+      "updated_at": "2026-06-29T05:03:10Z",
+      "scores": [
+        {
+          "dataset_name": "일반 상식",
+          "metric_type": "raw",
+          "score": 16.06
+        },
+        {
+          "dataset_name": "멀티 모달",
+          "metric_type": "raw",
+          "score": 31.43
+        },
+        {
+          "dataset_name": "수치 정보",
+          "metric_type": "raw",
+          "score": 9.52
+        },
+        {
+          "dataset_name": "팩트 체크",
+          "metric_type": "raw",
+          "score": 58.57
+        },
+        {
+          "dataset_name": "특수 도메인",
+          "metric_type": "raw",
+          "score": 100.0
+        },
+        {
+          "dataset_name": "K-Prism (Text, Intact)",
+          "metric_type": "raw",
+          "score": 56.94
+        },
+        {
+          "dataset_name": "K-Prism (Text, Perturbed)",
+          "metric_type": "raw",
+          "score": 55.16
+        },
+        {
+          "dataset_name": "K-Prism (Image, Intact)",
+          "metric_type": "raw",
+          "score": 79.32
+        },
+        {
+          "dataset_name": "K-Prism (Image, Perturbed)",
+          "metric_type": "raw",
+          "score": 44.58
+        },
+        {
+          "dataset_name": "소버린 벤치마크",
+          "metric_type": "raw",
+          "score": 31.51
+        }
+      ]
+    },
+    {
+      "provider": "SK Telecom",
+      "name": "ax4.0",
+      "repo": "skt/A.X-4.0-VL-Light",
+      "is_multimodal": true,
+      "status": "completed",
+      "updated_at": "2026-06-29T05:31:35Z",
+      "scores": [
+        {
+          "dataset_name": "일반 상식",
+          "metric_type": "raw",
+          "score": 20.73
+        },
+        {
+          "dataset_name": "멀티 모달",
+          "metric_type": "raw",
+          "score": 38.21
+        },
+        {
+          "dataset_name": "수치 정보",
+          "metric_type": "raw",
+          "score": 36.67
+        },
+        {
+          "dataset_name": "팩트 체크",
+          "metric_type": "raw",
+          "score": 74.29
+        },
+        {
+          "dataset_name": "특수 도메인",
+          "metric_type": "raw",
+          "score": 100.0
+        },
+        {
+          "dataset_name": "K-Prism (Text, Intact)",
+          "metric_type": "raw",
+          "score": 61.51
+        },
+        {
+          "dataset_name": "K-Prism (Text, Perturbed)",
+          "metric_type": "raw",
+          "score": 61.51
+        },
+        {
+          "dataset_name": "K-Prism (Image, Intact)",
+          "metric_type": "raw",
+          "score": 84.14
+        },
+        {
+          "dataset_name": "K-Prism (Image, Perturbed)",
+          "metric_type": "raw",
+          "score": 8.63
+        },
+        {
+          "dataset_name": "소버린 벤치마크",
+          "metric_type": "raw",
+          "score": 30.85
+        }
+      ]
+    },
+    {
+      "provider": "NAVER Cloud",
+      "name": "clova3b",
+      "repo": "naver-hyperclovax/HyperCLOVAX-SEED-Vision-Instruct-3B",
+      "is_multimodal": true,
+      "status": "completed",
+      "updated_at": "2026-06-29T05:44:09Z",
+      "scores": [
+        {
+          "dataset_name": "일반 상식",
+          "metric_type": "raw",
+          "score": 15.03
+        },
+        {
+          "dataset_name": "멀티 모달",
+          "metric_type": "raw",
+          "score": 26.43
+        },
+        {
+          "dataset_name": "수치 정보",
+          "metric_type": "raw",
+          "score": 38.57
+        },
+        {
+          "dataset_name": "팩트 체크",
+          "metric_type": "raw",
+          "score": 67.14
+        },
+        {
+          "dataset_name": "특수 도메인",
+          "metric_type": "raw",
+          "score": 100.0
+        },
+        {
+          "dataset_name": "K-Prism (Text, Intact)",
+          "metric_type": "raw",
+          "score": 8.33
+        },
+        {
+          "dataset_name": "K-Prism (Text, Perturbed)",
+          "metric_type": "raw",
+          "score": 74.8
+        },
+        {
+          "dataset_name": "K-Prism (Image, Intact)",
+          "metric_type": "raw",
+          "score": 67.07
+        },
+        {
+          "dataset_name": "K-Prism (Image, Perturbed)",
+          "metric_type": "raw",
+          "score": 63.05
+        },
+        {
+          "dataset_name": "소버린 벤치마크",
+          "metric_type": "raw",
+          "score": 13.13
+        }
+      ]
+    },
+    {
+      "provider": "OpenAI",
+      "name": "gpt-oss-20b",
+      "repo": "openai/gpt-oss-20b",
+      "is_multimodal": false,
+      "status": "completed",
+      "updated_at": "2026-06-29T06:09:40Z",
+      "scores": [
+        {
+          "dataset_name": "일반 상식",
+          "metric_type": "raw",
+          "score": 25.39
+        },
+        {
+          "dataset_name": "수치 정보",
+          "metric_type": "raw",
+          "score": 38.1
+        },
+        {
+          "dataset_name": "특수 도메인",
+          "metric_type": "raw",
+          "score": 100.0
+        },
+        {
+          "dataset_name": "소버린 벤치마크",
+          "metric_type": "raw",
+          "score": 15.1
+        }
+      ]
+    },
+    {
+      "provider": "Shanghai AI Laboratory (OpenGVLab)",
+      "name": "InternVL3-9B",
+      "repo": "OpenGVLab/InternVL3-9B",
+      "is_multimodal": true,
+      "status": "completed",
+      "updated_at": "2026-06-29T06:16:28Z",
+      "scores": [
+        {
+          "dataset_name": "일반 상식",
+          "metric_type": "raw",
+          "score": 41.97
+        },
+        {
+          "dataset_name": "멀티 모달",
+          "metric_type": "raw",
+          "score": 29.64
+        },
+        {
+          "dataset_name": "수치 정보",
+          "metric_type": "raw",
+          "score": 39.52
+        },
+        {
+          "dataset_name": "K-Prism (Text, Intact)",
+          "metric_type": "raw",
+          "score": 57.74
+        },
+        {
+          "dataset_name": "K-Prism (Text, Perturbed)",
+          "metric_type": "raw",
+          "score": 57.34
+        },
+        {
+          "dataset_name": "K-Prism (Image, Intact)",
+          "metric_type": "raw",
+          "score": 81.12
+        },
+        {
+          "dataset_name": "K-Prism (Image, Perturbed)",
+          "metric_type": "raw",
+          "score": 29.92
+        },
+        {
+          "dataset_name": "팩트 체크",
+          "metric_type": "raw",
+          "score": 67.14
+        },
+        {
+          "dataset_name": "소버린 벤치마크",
+          "metric_type": "raw",
+          "score": 38.07
+        }
+      ]
+    },
+    {
+      "provider": "Shanghai AI Laboratory (OpenGVLab)",
+      "name": "InternVL3-14B",
+      "repo": "OpenGVLab/InternVL3-14B",
+      "is_multimodal": true,
+      "status": "completed",
+      "updated_at": "2026-06-29T06:37:45Z",
+      "scores": [
+        {
+          "dataset_name": "일반 상식",
+          "metric_type": "raw",
+          "score": 52.85
+        },
+        {
+          "dataset_name": "멀티 모달",
+          "metric_type": "raw",
+          "score": 29.29
+        },
+        {
+          "dataset_name": "수치 정보",
+          "metric_type": "raw",
+          "score": 38.1
+        },
+        {
+          "dataset_name": "팩트 체크",
+          "metric_type": "raw",
+          "score": 71.43
+        },
+        {
+          "dataset_name": "K-Prism (Text, Intact)",
+          "metric_type": "raw",
+          "score": 61.31
+        },
+        {
+          "dataset_name": "K-Prism (Text, Perturbed)",
+          "metric_type": "raw",
+          "score": 66.47
+        },
+        {
+          "dataset_name": "K-Prism (Image, Intact)",
+          "metric_type": "raw",
+          "score": 78.51
+        },
+        {
+          "dataset_name": "K-Prism (Image, Perturbed)",
+          "metric_type": "raw",
+          "score": 27.31
+        },
+        {
+          "dataset_name": "소버린 벤치마크",
+          "metric_type": "raw",
+          "score": 47.26
+        }
+      ]
+    },
+    {
+      "provider": "Shanghai AI Laboratory (OpenGVLab)",
+      "name": "InternVL3-38B",
+      "repo": "OpenGVLab/InternVL3-38B",
+      "is_multimodal": true,
+      "status": "completed",
+      "updated_at": "2026-06-29T06:58:18Z",
+      "scores": [
+        {
+          "dataset_name": "일반 상식",
+          "metric_type": "raw",
+          "score": 56.99
+        },
+        {
+          "dataset_name": "멀티 모달",
+          "metric_type": "raw",
+          "score": 30.36
+        },
+        {
+          "dataset_name": "수치 정보",
+          "metric_type": "raw",
+          "score": 46.67
+        },
+        {
+          "dataset_name": "팩트 체크",
+          "metric_type": "raw",
+          "score": 75.71
+        },
+        {
+          "dataset_name": "K-Prism (Text, Intact)",
+          "metric_type": "raw",
+          "score": 61.11
+        },
+        {
+          "dataset_name": "K-Prism (Text, Perturbed)",
+          "metric_type": "raw",
+          "score": 59.92
+        },
+        {
+          "dataset_name": "K-Prism (Image, Intact)",
+          "metric_type": "raw",
+          "score": 78.11
+        },
+        {
+          "dataset_name": "K-Prism (Image, Perturbed)",
+          "metric_type": "raw",
+          "score": 39.96
+        },
+        {
+          "dataset_name": "소버린 벤치마크",
+          "metric_type": "raw",
+          "score": 56.89
+        }
+      ]
+    },
+    {
+      "provider": "Shanghai AI Laboratory (OpenGVLab)",
+      "name": "InternVL3-8B-Pretrained",
+      "repo": "OpenGVLab/InternVL3-8B-Pretrained",
+      "is_multimodal": true,
+      "status": "completed",
+      "updated_at": "2026-06-29T07:04:44Z",
+      "scores": [
+        {
+          "dataset_name": "일반 상식",
+          "metric_type": "raw",
+          "score": 41.45
+        },
+        {
+          "dataset_name": "멀티 모달",
+          "metric_type": "raw",
+          "score": 17.5
+        },
+        {
+          "dataset_name": "수치 정보",
+          "metric_type": "raw",
+          "score": 42.86
+        },
+        {
+          "dataset_name": "팩트 체크",
+          "metric_type": "raw",
+          "score": 60.0
+        },
+        {
+          "dataset_name": "K-Prism (Text, Intact)",
+          "metric_type": "raw",
+          "score": 58.33
+        },
+        {
+          "dataset_name": "K-Prism (Text, Perturbed)",
+          "metric_type": "raw",
+          "score": 57.14
+        },
+        {
+          "dataset_name": "K-Prism (Image, Intact)",
+          "metric_type": "raw",
+          "score": 65.46
+        },
+        {
+          "dataset_name": "K-Prism (Image, Perturbed)",
+          "metric_type": "raw",
+          "score": 47.59
+        },
+        {
+          "dataset_name": "소버린 벤치마크",
+          "metric_type": "raw",
+          "score": 45.95
+        }
+      ]
+    },
+    {
+      "provider": "Shanghai AI Laboratory (OpenGVLab)",
+      "name": "InternVL3-14B-Pretrained",
+      "repo": "OpenGVLab/InternVL3-14B-Pretrained",
+      "is_multimodal": true,
+      "status": "completed",
+      "updated_at": "2026-06-29T07:12:43Z",
+      "scores": [
+        {
+          "dataset_name": "일반 상식",
+          "metric_type": "raw",
+          "score": 54.4
+        },
+        {
+          "dataset_name": "멀티 모달",
+          "metric_type": "raw",
+          "score": 25.71
+        },
+        {
+          "dataset_name": "수치 정보",
+          "metric_type": "raw",
+          "score": 42.38
+        },
+        {
+          "dataset_name": "팩트 체크",
+          "metric_type": "raw",
+          "score": 70.0
+        },
+        {
+          "dataset_name": "K-Prism (Text, Intact)",
+          "metric_type": "raw",
+          "score": 63.1
+        },
+        {
+          "dataset_name": "K-Prism (Text, Perturbed)",
+          "metric_type": "raw",
+          "score": 72.22
+        },
+        {
+          "dataset_name": "K-Prism (Image, Intact)",
+          "metric_type": "raw",
+          "score": 78.11
+        },
+        {
+          "dataset_name": "K-Prism (Image, Perturbed)",
+          "metric_type": "raw",
+          "score": 54.02
+        },
+        {
+          "dataset_name": "소버린 벤치마크",
+          "metric_type": "raw",
+          "score": 50.33
+        }
+      ]
+    }
+  ]
+}

front/leaderboard.mjs

@@ -0,0 +1,413 @@
+import { config } from './config.mjs';
+import { loadBenchmark } from './data.mjs';
+
+// ===== Config =====
+const MAX_COMPARE = 999; // Allow selecting all models
+
+// ===== State =====
+const state = {
+  datasets: [],
+  models: [],
+  selectedDatasets: new Set(),
+  providerFilter: "",
+  search: "",
+  compare: new Set(),
+  dark: window.matchMedia && window.matchMedia('(prefers-color-scheme: dark)').matches,
+};
+
+// ===== Utils =====
+const $ = (s)=>document.querySelector(s);
+const $$ = (s)=>Array.from(document.querySelectorAll(s));
+const showToast=(m)=>{const t=$("#toast");$("#toastMsg").textContent=m;t.classList.remove("hidden");setTimeout(()=>t.classList.add("hidden"),2200);};
+const fmt=(n,d=2)=>(n==null||Number.isNaN(n))?"—":Number(n).toFixed(d);
+const escapeHTML = value => String(value).replace(/[&<>"']/g, char => ({
+  '&': '&amp;', '<': '&lt;', '>': '&gt;', '"': '&quot;', "'": '&#39;',
+})[char]);
+const slug=(s)=>s.toLowerCase().replace(/[^a-z0-9]+/g,'-');
+const TEXT_ONLY_EXCLUDED_DATASETS = new Set([
+  "멀티 모달", "팩트 체크",
+  "K-Prism (Text, Intact)", "K-Prism (Text, Perturbed)",
+  "K-Prism (Image, Intact)", "K-Prism (Image, Perturbed)"
+]);
+
+function isUnsupportedForTextOnly(model, dataset){
+  return !model.is_multimodal && TEXT_ONLY_EXCLUDED_DATASETS.has(dataset);
+}
+
+let hasDatasetPreference = false;
+let hasLoadedData = false;
+let isLoading = false;
+
+function savePrefs(){
+  try {
+    localStorage.setItem('llm_lb_prefs', JSON.stringify({
+      selectedDatasets: [...state.selectedDatasets],
+      providerFilter: state.providerFilter,
+      search: state.search, dark: state.dark,
+      compare: [...state.compare],
+    }));
+  } catch { /* Browsers may block storage inside embedded Spaces. */ }
+}
+function loadPrefs(){
+  try {
+    const p = JSON.parse(localStorage.getItem('llm_lb_prefs') || '{}');
+    if (Array.isArray(p.selectedDatasets)) {
+      state.selectedDatasets = new Set(p.selectedDatasets.filter(x => typeof x === 'string'));
+      hasDatasetPreference = true;
+    }
+    if (typeof p.providerFilter === 'string') state.providerFilter = p.providerFilter;
+    if (typeof p.search === 'string') state.search = p.search;
+    if (typeof p.dark === 'boolean') state.dark = p.dark;
+    if (Array.isArray(p.compare)) state.compare = new Set(p.compare.filter(x => typeof x === 'string'));
+  } catch { /* Preferences are optional. */ }
+}
+
+// ===== Rendering =====
+function visibleDatasets(){return state.selectedDatasets.size ? state.datasets.filter(d=>state.selectedDatasets.has(d)) : []; }
+
+function modelRowMean(model, dsList){
+  // 텍스트 전용 모델이면 지원하지 않는 데이터셋은 평균 계산에서 제외
+  const effDs = dsList.filter(d => !isUnsupportedForTextOnly(model, d));
+  const map=Object.fromEntries(model.scores.map(s=>[s.dataset_name,s]));
+  const vals=effDs.map(d=>map[d]?.score).filter(v=>typeof v==='number');
+  return (vals.length >= effDs.length && effDs.length>0) ? vals.reduce((a,b)=>a+b,0)/vals.length : null;
+}
+function renderDatasetFilters(){
+  const c=$("#datasetFilters"); c.innerHTML='';
+  state.datasets.forEach(d=>{
+    const checked=state.selectedDatasets.has(d);
+    c.insertAdjacentHTML('beforeend', `
+      <label class="flex flex-nowrap items-center gap-2 rounded-lg border" style="border-color: var(--border); padding:.5rem .6rem;">
+        <input type="checkbox" class="h-4 w-4 accent-indigo-600" data-dataset="${escapeHTML(d)}" ${checked?'checked':''}/>
+        <span class="text-sm" style="color: var(--text)">${escapeHTML(d)}</span>
+      </label>
+    `);
+  });
+  c.querySelectorAll('input[type="checkbox"]').forEach(cb=>{
+    cb.addEventListener('change', ()=>{
+      if(cb.checked) state.selectedDatasets.add(cb.dataset.dataset);
+      else state.selectedDatasets.delete(cb.dataset.dataset);
+      if([...c.querySelectorAll('input:checked')].length===0) state.selectedDatasets=new Set();
+      savePrefs(); renderTable(); updateChart();
+    });
+  });
+  $("#kpiDatasets").textContent = state.datasets.length;
+}
+
+function renderProviders(){
+  const provs=[...new Set(state.models.map(m=>m.provider))].sort();
+  const sel=$("#providerSelect");
+  sel.replaceChildren(new Option('All providers', ''));
+  provs.forEach(provider => sel.add(new Option(provider, provider)));
+  sel.value = state.providerFilter;
+}
+
+function renderHeader(){
+  const head=$("#tableHeaderRow");
+  head.querySelectorAll('th[data-ds]').forEach(el=>el.remove());
+  visibleDatasets().forEach(d=>{
+    const th=document.createElement('th');
+    th.dataset.ds=d; th.className='p-3 text-left ds-col ds-head';
+    th.innerHTML = `<div class="font-semibold" style="color: var(--text)">${escapeHTML(d)}</div>`;
+    head.appendChild(th);
+  });
+}
+
+function renderTable(){
+  renderHeader();
+  const ds=visibleDatasets();
+  const tb=$("#tableBody"); tb.innerHTML='';
+
+  let models=state.models.filter(m=>{
+    const text=(m.provider+" "+m.name+" "+m.repo).toLowerCase();
+    const okProvider=!state.providerFilter || m.provider===state.providerFilter;
+    const okSearch=!state.search || text.includes(state.search.toLowerCase());
+    return okProvider && okSearch;
+  }).map(m=>({...m, mean:modelRowMean(m, ds)}))
+    .sort((a,b)=>(b.mean??-1)-(a.mean??-1));
+
+  models.forEach((m, i)=>{
+    const sMap=Object.fromEntries(m.scores.map(s=>[s.dataset_name, s]));
+    const id=slug(m.provider+"-"+m.name);
+    const favOn=state.compare.has(id);
+
+    const row=document.createElement('tr');
+    row.className='border-b'; row.style.borderColor=getComputedStyle(document.documentElement).getPropertyValue('--border');
+    row.innerHTML = `
+      <td class="p-3 rank-col sticky-rank">${i+1}</td>
+      <td class="p-3 provider-col sticky-provider border-l">
+        <div class="font-medium">${escapeHTML(m.provider)}</div>
+      </td>
+      <td class="p-3 model-col sticky-model border-l">
+        <div class="flex items-center gap-2">
+          <button class="text-lg" data-fav="${id}" title="Add to compare">${favOn?'⭐':'☆'}</button>
+          <div class="font-semibold" style="color: var(--text)">${escapeHTML(m.name)}</div>
+        </div>
+        <div class="text-xs text-slate-500">${escapeHTML(m.repo)}</div>
+        <div class="text-xs text-slate-500">${m.is_multimodal ? 'Multimodal' : 'Text-only'}</div>
+      </td>
+      <td class="p-3 mean-col sticky-mean border-l">
+        <div class="text-base font-semibold text-blue-500">
+          ${m.mean == null ? '<span class="text-slate-400" title="No complete score for the selected datasets">—</span>' : fmt(m.mean,2)}
+        </div>
+        <!--div class="text-[11px] text-slate-500">mean across selected</div-->
+      </td>
+    `;
+    ds.forEach(d=>{
+      const s=sMap[d]; const mt=s?.metric_type;
+      // const badge= mt==='llm-as-judge' ? 'metric-badge-judge' : 'metric-badge';
+      const cell=document.createElement('td');
+      cell.className='p-3 ds-col'; cell.title='';
+      // 텍스트 전용 모델이면 지원하지 않는 데이터셋 컬럼은 항상 비움(—)
+      if (isUnsupportedForTextOnly(m, d)) {
+        cell.innerHTML = '<span class="text-slate-300">—</span>';
+        row.appendChild(cell);
+        return;
+      }
+
+      {
+          cell.innerHTML = s ? (
+            mt==='llm-as-judge' ? `
+                <div class="flex item-center">
+                    <div class="text-base font-semibold" style="color: var(--text)">${fmt(s.score,2)}</div>
+                    <div class="pill metric-badge-judge ms-2">${mt}</div>
+                </div>
+            ` :
+            `
+                <div class="flex item-center">
+                    <div class="text-base font-semibold" style="color: var(--text)">${fmt(s.score,2)}</div>
+                </div>
+            `
+          ) : '<span class="text-slate-300">—</span>';
+      }
+      row.appendChild(cell);
+    });
+    tb.appendChild(row);
+  });
+
+  if (!models.length) {
+    const row = document.createElement('tr');
+    const cell = document.createElement('td');
+    cell.colSpan = 4 + ds.length;
+    cell.className = 'p-6 text-center text-slate-500';
+    cell.textContent = !hasLoadedData ? 'Results are unavailable.'
+      : state.models.length ? 'No models match your filters.' : 'No results have been published yet.';
+    row.appendChild(cell);
+    tb.appendChild(row);
+  }
+  $("#kpiModels").textContent = hasLoadedData ? models.length : '—';
+
+  $$('button[data-fav]').forEach(b=>{
+    b.onclick=()=>{
+      const id=b.dataset.fav;
+      if(state.compare.has(id)) state.compare.delete(id);
+      else{
+        if(state.compare.size>=MAX_COMPARE){ showToast(`You can compare up to ${MAX_COMPARE} models.`); return; }
+        state.compare.add(id);
+      }
+      savePrefs(); renderTable(); updateChart();
+    };
+  });
+}
+
+// ===== Chart =====
+let chart;
+const baseColors = [
+  '#4f46e5', '#10b981', '#f59e0b', '#ef4444', '#8b5cf6', '#ec4899', '#06b6d4', '#84cc16', '#14b8a6', '#f43f5e'
+];
+function updateChart(){
+  const chartUnavailable = typeof window.Chart !== 'function';
+  $("#chartUnavailable").classList.toggle('hidden', !chartUnavailable);
+  $("#scoresChart").classList.toggle('hidden', chartUnavailable);
+  if (chartUnavailable) return;
+  const ds=visibleDatasets();
+  const labels=ds;
+  const datasets=[];
+  const chosen=state.models.filter(m=>state.compare.has(slug(m.provider+"-"+m.name)));
+  chosen.forEach((m, idx)=>{
+    const map=Object.fromEntries(m.scores.map(s=>[s.dataset_name, s.score]));
+    const color = baseColors[idx % baseColors.length];
+
+    const lineData = labels.map(d => {
+      if (isUnsupportedForTextOnly(m, d)) return null;
+      if (d.startsWith("K-Prism")) return null;
+      return map[d] ?? null;
+    });
+
+    const barData = labels.map(d => {
+      if (isUnsupportedForTextOnly(m, d)) return null;
+      if (!d.startsWith("K-Prism")) return null;
+      return map[d] ?? null;
+    });
+
+    // Add line dataset
+    datasets.push({
+      type: 'line',
+      label: `${m.provider} / ${m.name}`,
+      data: lineData,
+      borderColor: color,
+      backgroundColor: color,
+      tension: .25, spanGaps: true, borderWidth: 3, pointRadius: 4, pointHoverRadius: 6,
+    });
+
+    // Add bar dataset
+    datasets.push({
+      type: 'bar',
+      label: `${m.provider} / ${m.name} (Bar)`,
+      data: barData,
+      borderColor: color,
+      backgroundColor: color + '80', // Add transparency for bars
+      borderWidth: 2,
+      borderRadius: 4,
+    });
+  });
+
+  const ctx=document.getElementById('scoresChart').getContext('2d');
+  if(!chart){
+    chart=new Chart(ctx,{
+      data:{labels, datasets},
+      options:{
+        responsive:true,
+        maintainAspectRatio:false,  // uses #chartWrap height (clamp -> responsive)
+        plugins:{
+          legend:{
+            display:true,
+            labels:{
+              boxWidth:18,
+              usePointStyle:true,
+              filter: function(item, chart) {
+                // Hide the bar datasets from the legend to avoid duplicates
+                return !item.text.includes('(Bar)');
+              }
+            }
+          },
+          tooltip:{ mode:'index', intersect:false }
+        },
+        interaction:{ mode:'index', intersect:false },
+        scales:{
+          y:{ beginAtZero:true, max:100, title:{display:true, text:'Score'}, grid:{ drawBorder:false }},
+          x:{
+            type: 'category',
+            offset:true,
+            grid:{ display:false },
+            ticks:{ autoSkip:false, maxRotation:45, minRotation:45 }
+          }
+        },
+        layout:{ padding:0 }
+      }
+    });
+  }else{
+    chart.data.labels=labels;
+    chart.data.datasets=datasets;
+    chart.update();
+  }
+}
+
+// ===== Tabs / Theme / Export =====
+function bindTabs(){
+  $$(".tab").forEach(btn=>{
+    btn.addEventListener('click', ()=>{
+      const t=btn.dataset.tab;
+      $$(".tab").forEach(b=>b.classList.remove('tab-active')); btn.classList.add('tab-active');
+      ["leaderboard","about"].forEach(x=>{
+        const el=document.getElementById(`tab-${x}`); (x===t)?el.classList.remove('hidden'):el.classList.add('hidden');
+      });
+    });
+  });
+}
+function applyTheme(){
+    document.documentElement.classList.toggle('dark', state.dark);
+    $$('[data-action="theme"]').forEach(button => {
+      button.textContent = state.dark ? '☀️' : '🌙';
+    });
+}
+
+function exportCSV(){
+  const ds=visibleDatasets();
+  const headers=['Rank','Provider','Model','Repo','Mean',...ds];
+  const rows=[];
+  let models=state.models.map(m=>({...m, mean:modelRowMean(m, ds)})).sort((a,b)=>(b.mean??-1)-(a.mean??-1));
+  models=models.filter(m=>{
+    const text=(m.provider+" "+m.name+" "+m.repo).toLowerCase();
+    const okProvider=!state.providerFilter || m.provider===state.providerFilter;
+    const okSearch=!state.search || text.includes(state.search.toLowerCase());
+    return okProvider && okSearch;
+  });
+  models.forEach((m,i)=>{
+    const map=Object.fromEntries(m.scores.map(s=>[s.dataset_name, s]));
+    const row=[i+1, m.provider, m.name, m.repo, fmt(m.mean,2)];
+    ds.forEach(d=>{
+      if (isUnsupportedForTextOnly(m, d)) {
+        row.push('');
+        return;
+      }
+      const s=map[d];
+      row.push(s ? `${fmt(s.score,2)} (${s.metric_type})` : '');
+    });
+    rows.push(row);
+  });
+  const csv=[headers, ...rows].map(r=> r.map(x=>'"'+String(x).replaceAll('"','""')+'"').join(',')).join('\n');
+  const blob=new Blob([csv],{type:'text/csv;charset=utf-8;'}); const url=URL.createObjectURL(blob);
+  const a=document.createElement('a'); a.href=url; a.download='keti-leaderboard.csv'; a.click(); URL.revokeObjectURL(url);
+}
+
+// ===== Controls =====
+function bindControls(){
+  $("#selectAllBtn").onclick=()=>{ state.selectedDatasets=new Set(state.datasets); renderDatasetFilters(); renderTable(); updateChart(); savePrefs(); };
+  $("#clearAllBtn").onclick=()=>{ state.selectedDatasets=new Set(); renderDatasetFilters(); renderTable(); updateChart(); savePrefs(); };
+  $("#searchInput").addEventListener('input', (e)=>{ state.search=e.target.value; savePrefs(); renderTable(); });
+  $("#providerSelect").addEventListener('change', (e)=>{ state.providerFilter=e.target.value; savePrefs(); renderTable(); });
+  $$('[data-action="refresh"]').forEach(button => button.onclick = refreshResults);
+  $("#exportCsvBtn").onclick=exportCSV;
+  $$('[data-action="theme"]').forEach(button => {
+    button.onclick=()=>{ state.dark=!state.dark; applyTheme(); savePrefs(); };
+  });
+}
+
+async function refreshResults(){
+  if (isLoading) return;
+  isLoading = true;
+  const status = $("#dataStatus");
+  status.dataset.state = 'loading';
+  status.textContent = 'Loading published results…';
+  $$('[data-action="refresh"]').forEach(button => button.disabled = true);
+  try {
+    const result = await loadBenchmark(config.dataUrl, config.requestTimeoutMs);
+    state.datasets = result.datasets;
+    state.models = result.models;
+    state.selectedDatasets = hasDatasetPreference
+      ? new Set(state.datasets.filter(dataset => state.selectedDatasets.has(dataset)))
+      : new Set(state.datasets);
+    hasDatasetPreference = true;
+    hasLoadedData = true;
+    const modelIds = new Set(state.models.map(model => slug(model.provider + '-' + model.name)));
+    state.compare = new Set([...state.compare].filter(id => modelIds.has(id)));
+    if (!state.models.some(model => model.provider === state.providerFilter)) state.providerFilter = '';
+    const timestamps = state.models.map(model => Date.parse(model.updated_at)).filter(Number.isFinite);
+    $("#lastUpdated").textContent = timestamps.length
+      ? 'Last updated ' + new Date(Math.max(...timestamps)).toLocaleString() : 'Last updated —';
+    renderDatasetFilters(); renderProviders(); renderTable(); updateChart();
+    $("#exportCsvBtn").disabled = !state.models.length;
+    status.dataset.state = 'ready';
+    status.textContent = `${state.models.length} models · ${state.datasets.length} datasets · Published results`;
+    savePrefs();
+  } catch (error) {
+    console.error('Unable to load published results:', error);
+    status.dataset.state = 'error';
+    status.textContent = hasLoadedData
+      ? 'Could not refresh results. Previously loaded results are still shown. Try Refresh again.'
+      : 'Could not load results. Please check your connection and try Refresh again.';
+    if (!hasLoadedData) {
+      renderTable();
+      $("#kpiDatasets").textContent = '—';
+    }
+  } finally {
+    isLoading = false;
+    $$('[data-action="refresh"]').forEach(button => button.disabled = false);
+  }
+}
+
+// Bind once: refreshing must not register duplicate listeners.
+loadPrefs(); applyTheme(); bindTabs(); bindControls();
+$("#searchInput").value = state.search;
+refreshResults();

index.html

@@ -0,0 +1,246 @@
+<!DOCTYPE html>
+<html lang="en">
+<head>
+  <meta charset="UTF-8" />
+  <meta name="viewport" content="width=device-width, initial-scale=1.0"/>
+  <meta name="author" content="Dr. Mohamed Sana">
+  <title>KETI LLM Benchmarks</title>
+
+  <!-- Tailwind + Chart.js -->
+  <script src="https://cdn.tailwindcss.com"></script>
+  <script src="https://cdn.jsdelivr.net/npm/chart.js"></script>
+
+  <!-- Fonts -->
+  <link rel="preconnect" href="https://fonts.googleapis.com">
+  <link rel="preconnect" href="https://fonts.gstatic.com" crossorigin>
+  <link href="https://fonts.googleapis.com/css2?family=Inter:wght@300;400;500;600;700;800&display=swap" rel="stylesheet">
+
+  <style>
+    :root{
+      color-scheme: light dark;
+      --text: #0f172a;
+      --bg: #f8fafc;
+      --card-bg: rgba(255,255,255,.75);
+      --border: #e2e8f0;
+      --header-bg: rgba(241,245,249,.85);
+      --sticky-bg: rgba(248,250,252,.92);
+      --chip-bg:#f1f5f9;
+    }
+    .dark{
+      --text: #e2e8f0;
+      --bg: #020617;
+      --card-bg: rgba(2,6,23,.6);
+      --border:#334155;
+      --header-bg: rgba(30,41,59,.75);
+      --sticky-bg: rgba(2,6,23,.92);
+      --chip-bg: rgba(15,23,42,.6);
+    }
+    html { font-family: Inter, system-ui, -apple-system, Segoe UI, Roboto, "Helvetica Neue", Arial, "Noto Sans"; }
+    body { color: var(--text); background: var(--bg); }
+
+    .card{ border-radius:1rem; box-shadow:0 10px 25px rgba(2,6,23,.08); background:var(--card-bg); backdrop-filter: blur(8px); border:1px solid var(--border); }
+    .btn{ display:inline-flex; align-items:center; justify-content:center; gap:.5rem; border-radius:.8rem; padding:.625rem 1rem; font-weight:700; transition: transform .05s ease; }
+    .btn:disabled{ opacity:.5; cursor:wait; }
+    #dataStatus[data-state="error"]{ border-color:#f59e0b; }
+    .btn:active{ transform: scale(.98); }
+    .btn-primary{ background:#4f46e5; color:#fff; } .btn-primary:hover{ background:#6366f1; }
+    .btn-ghost{ background:transparent; border:1px solid var(--border); } .btn-ghost:hover{ background:var(--chip-bg); }
+    .btn-outline{ border:1px solid #4f46e5; color:#4338ca; } .dark .btn-outline{ color:#a5b4fc; }
+    .btn-outline:hover{ background:#eef2ff; } .dark .btn-outline:hover{ background: rgba(30,27,75,.5); }
+    .input{ width:100%; border:1px solid var(--border); border-radius:.8rem; padding:.6rem .8rem; background:#fff; color:#0f172a; }
+    .dark .input{ background:#0b1220; color:#e2e8f0; }
+    .input:focus{ outline:none; box-shadow:0 0 0 2px rgba(99,102,241,.6); }
+    .label{ font-size:.875rem; font-weight:600; color:#334155; } .dark .label{ color:#cbd5e1; }
+    .tab{ padding:.6rem 1rem; border-radius:.8rem; cursor:pointer; font-weight:700; }
+    .tab-active{ background:#fff; border:1px solid var(--border); box-shadow:0 1px 2px rgba(0,0,0,.04); } .dark .tab-active{ background:#0b1220; }
+    .pill{ display:inline-flex; align-items:center; padding:.1rem .5rem; border-radius:999px; font-size:.53rem; font-weight:700; }
+    .metric-badge{ background:#ecfdf5; color:#065f46; } .dark .metric-badge{ background:rgba(16,185,129,.18); color:#d1fae5; }
+    .metric-badge-judge{ background:#e0f2fe; color:#075985; } .dark .metric-badge-judge{ background:rgba(56,189,248,.18); color:#bae6fd; }
+    .kpi{ color:#475569; font-size:.875rem; } .dark .kpi{ color:#94a3b8; }
+
+    /* Column widths */
+    /* .rank-col{ width:1.5rem; }
+    .provider-col{ width:4rem;  }
+    .model-col{ width:10rem; white-space:nowrap; }
+    .mean-col{ width:3rem; white-space:nowrap; }
+    .ds-col{ min-width:9.5rem; white-space:nowrap; } */
+
+    .rank-col{ }
+    .provider-col{   }
+    .model-col{ white-space:nowrap; }
+    .mean-col{ white-space:nowrap;  }
+    .ds-col{ white-space:nowrap; }
+
+    /* Sticky columns: use calc so offsets update easily */
+    .sticky-rank{ position: sticky; left: 0; z-index: 3; background: var(--sticky-bg); }
+    .sticky-provider{ position: sticky; left: 2.3rem; z-index: 3; background: var(--sticky-bg); }
+    .sticky-model{ position: sticky; left: calc(1.85rem + 6rem); z-index: 3; background: var(--sticky-bg); }
+    .sticky-mean{ position: sticky; left: calc(2.3rem + 4rem + 10rem + 0.5rem); z-index: 3; background: var(--sticky-bg); }
+
+    /* Keep provider, model and mean visible together on narrow screens. */
+    @media (max-width: 768px){
+      .rank-col, .ds-col, .ds-head { display:none !important; }
+      table{ min-width:0 !important; table-layout:fixed; }
+      .provider-col{ width:28%; }
+      .model-col{ width:50%; white-space:normal; overflow-wrap:anywhere; }
+      .mean-col{ width:22%; }
+      .sticky-provider, .sticky-model, .sticky-mean{ position:static; }
+      td.provider-col{ overflow-wrap:anywhere; }
+      th, td{ padding:.5rem !important; }
+    }
+
+    /* Header background */
+    thead th{ background: var(--header-bg); }
+
+    /* Global gradient */
+    .gradient-bg{
+      background:
+        radial-gradient(1200px 600px at 20% -10%, rgba(99,102,241,.25), rgba(99,102,241,0) 60%),
+        radial-gradient(1200px 600px at 80% -10%, rgba(20,184,166,.2), rgba(20,184,166,0) 60%);
+    }
+
+    /* Scrollbar for horizontal overflow */
+    .scrollbar-thin::-webkit-scrollbar{ height:10px; }
+    .scrollbar-thin::-webkit-scrollbar-thumb{ background:#c7d2fe; border-radius:999px; }
+    .scrollbar-thin::-webkit-scrollbar-track{ background:transparent; }
+
+    /* Chart container: responsive height */
+    #chartWrap{ height: clamp(260px, 42vh, 460px); }
+  </style>
+</head>
+<body class="min-h-screen dark:bg-gray-800">
+  <!-- Header -->
+  <header class="sticky top-0 z-40 backdrop-blur dark:bg-white-800 border-b" style="border-color: var(--border)">
+    <div class="mx-auto max-w-7xl px-4 md:px-6 py-3 md:py-4 flex items-center justify-between">
+      <div class="flex items-center gap-3 md:gap-4">
+        <div class="h-10 w-10 rounded-xl bg-indigo-600 text-white grid place-items-center shadow-lg">📡</div>
+        <div>
+          <h1 class="text-lg md:text-xl font-extrabold tracking-tight" style="color: var(--text)">KETI LLM Benchmarks</h1>
+          <p class="text-xs md:text-sm">Benchmarking models across KETI datasets</p>
+        </div>
+      </div>
+      <div class="md:flex items-center gap-2 hidden">
+        <button id="refreshBtn" data-action="refresh" class="btn btn-ghost" title="Refresh results">⟲ Refresh</button>
+        <button id="themeToggle" data-action="theme" class="btn btn-ghost" title="Toggle dark mode">🌙</button>
+      </div>
+    </div>
+  </header>
+
+  <!-- Main -->
+  <main class="mx-auto max-w-7xl px-4 md:px-6 py-6 md:py-8 space-y-6 md:space-y-8">
+    <!-- Tabs -->
+    <div class="flex gap-2 md:gap-3">
+      <button class="tab tab-active" data-tab="leaderboard">🏆 Leaderboard</button>
+      <button class="tab" data-tab="about">ℹ️ About</button>
+    </div>
+
+    <div id="dataStatus" class="card p-4 text-sm" role="status" aria-live="polite" data-state="loading">Loading published results…</div>
+    <noscript><p class="card p-4">Enable JavaScript to view and filter the leaderboard.</p></noscript>
+
+    <!-- Leaderboard Tab -->
+    <section id="tab-leaderboard" class="space-y-6 md:space-y-8">
+      <!-- Controls -->
+      <div class="grid lg:grid-cols-4 gap-4 md:gap-6">
+        <div class="card p-4 md:p-6 lg:col-span-2">
+          <div class="flex items-center justify-between">
+            <h2 class="font-semibold text-base md:text-lg" style="color: var(--text)">Select Datasets</h2>
+            <div class="flex gap-2">
+              <button id="selectAllBtn" class="btn btn-ghost text-xs md:text-sm px-3">Select all</button>
+              <button id="clearAllBtn" class="btn btn-ghost text-xs md:text-sm px-3">Clear</button>
+            </div>
+          </div>
+          <div id="datasetFilters" class="mt-3 md:mt-4 flex flex-wrap  gap-2.5"></div> <!--grid-cols-1 sm:grid-cols-2 lg:grid-cols-3-->
+        </div>
+
+        <div class="card p-4 md:p-6">
+          <h2 class="font-semibold text-base md:text-lg" style="color: var(--text)">Search & Filter</h2>
+          <div class="mt-2 md:mt-3 space-y-2.5">
+            <input id="searchInput" class="input" placeholder="Search provider/model…"/>
+            <select id="providerSelect" class="input">
+              <option value="">All providers</option>
+            </select>
+          </div>
+        </div>
+
+        <div class="card p-4 md:p-6">
+          <h2 class="font-semibold text-base md:text-lg" style="color: var(--text)">KPI</h2>
+          <div class="mt-2 grid grid-cols-2 gap-3">
+            <div>
+              <div class="text-2xl md:text-3xl font-extrabold" id="kpiModels">—</div>
+              <div class="kpi">Models</div>
+            </div>
+            <div>
+              <div class="text-2xl md:text-3xl font-extrabold" id="kpiDatasets">—</div>
+              <div class="kpi">Datasets</div>
+            </div>
+            <div class="col-span-2 text-xs text-slate-500 dark:text-slate-400" id="lastUpdated">Last updated —</div>
+          </div>
+        </div>
+      </div>
+
+      <!-- Visualization -->
+      <div class="card p-5 md:p-6 lg:p-8">
+        <div class="flex items-center justify-between gap-3">
+          <h2 class="font-semibold text-base md:text-lg" style="color: var(--text)">Model Comparison</h2>
+          <div class="text-xs md:text-sm text-slate-500">Click ☆ on rows to compare models</div>
+        </div>
+        <div id="chartWrap" class="mt-3 md:mt-4">
+          <p id="chartUnavailable" class="hidden text-sm text-slate-500">The comparison chart could not be loaded. Scores are available in the table below.</p>
+          <canvas id="scoresChart"></canvas>
+        </div>
+      </div>
+
+      <!-- Table (always horizontally scrollable) -->
+      <div class="card overflow-hidden">
+        <div class="px-4 md:px-6 py-3 md:py-4 flex items-center justify-between">
+          <div class="font-semibold" style="color: var(--text)">Leaderboard</div>
+          <div class="text-xs md:text-sm text-slate-500">Sorted by mean score across selected datasets</div>
+        </div>
+
+        <div class="overflow-x-auto overscroll-x-contain scrollbar-thin pb-2">
+          <table class="min-w-[64rem] w-full text-sm">
+            <thead class="border-t border-b" style="border-color: var(--border)">
+              <tr id="tableHeaderRow">
+                <th class="p-3 text-left rank-col sticky-rank">#</th>
+                <th class="p-3 text-left provider-col sticky-provider border-l">Provider</th>
+                <th class="p-3 text-left model-col sticky-model border-l">Model</th>
+                <th class="p-3 text-left mean-col sticky-mean border-l">
+                    Mean
+                    <div class="text-[11px] text-slate-500">on selected</div>
+                </th>
+                <!-- Dataset columns injected here -->
+              </tr>
+            </thead>
+            <tbody id="tableBody"></tbody>
+          </table>
+        </div>
+
+        <div class="flex items-center gap-2 justify-end p-3">
+          <button id="refreshMobileBtn" data-action="refresh" class="btn btn-ghost md:hidden" title="Refresh results">⟲ Refresh</button>
+          <button id="exportCsvBtn" disabled class="btn btn-outline" title="Export current view to CSV">⭳ Export CSV</button>
+          <button id="themeToggleMobile" data-action="theme" class="btn btn-ghost md:hidden" title="Toggle dark mode">🌙</button>
+        </div>
+      </div>
+    </section>
+
+    <!-- About Tab -->
+    <section id="tab-about" class="hidden">
+      <div class="card p-5 md:p-6 lg:p-8 space-y-3">
+        <h2 class="text-lg md:text-xl font-bold" style="color: var(--text)">About this Leaderboard</h2>
+        <p class="text-sm text-slate-600 dark:text-slate-300">
+          This leaderboard presents published model evaluations on KETI's ethicality and veracity benchmark datasets.
+        </p>
+        <p class="text-sm text-slate-600 dark:text-slate-300">
+          Model submissions and automatic evaluations are not available on this page. Results are updated when the maintainers publish a new evaluation.
+        </p>
+      </div>
+    </section>
+  </main>
+
+  <div id="toast" class="fixed bottom-4 left-1/2 -translate-x-1/2 hidden">
+    <div class="rounded-xl bg-slate-900 text-white px-4 py-2 shadow-lg">✅ <span id="toastMsg">Done</span></div>
+  </div>
+
+  <script type="module" src="./front/leaderboard.mjs"></script>
+</body>
+</html>

tests/test_static_space.py

@@ -0,0 +1,204 @@
+"""Browser regression checks for the static Space.
+
+Install playwright and Chrome/Chromium, then run:
+    python3 -m unittest discover -s tests -v
+Set CHROME_PATH to override the installed browser executable.
+"""
+import csv
+import io
+import json
+import os
+from functools import partial
+from http.server import SimpleHTTPRequestHandler, ThreadingHTTPServer
+from pathlib import Path
+import shutil
+import threading
+import unittest
+
+from playwright.sync_api import sync_playwright
+
+ROOT = Path(__file__).resolve().parents[1]
+SNAPSHOT = json.loads((ROOT / 'front/data/benchmark.json').read_text())
+
+
+class QuietHandler(SimpleHTTPRequestHandler):
+    def log_message(self, *args):
+        pass
+
+
+class StaticSpaceTests(unittest.TestCase):
+    @classmethod
+    def setUpClass(cls):
+        cls.server = ThreadingHTTPServer(
+            ('127.0.0.1', 0), partial(QuietHandler, directory=str(ROOT)))
+        cls.thread = threading.Thread(target=cls.server.serve_forever, daemon=True)
+        cls.thread.start()
+        cls.base = 'http://127.0.0.1:' + str(cls.server.server_port)
+        cls.playwright = sync_playwright().start()
+        cls.browser = cls.playwright.chromium.launch(
+            executable_path=os.environ.get('CHROME_PATH') or shutil.which('google-chrome')
+            or shutil.which('chromium'), headless=True, timeout=20000)
+
+    @classmethod
+    def tearDownClass(cls):
+        cls.browser.close()
+        cls.playwright.stop()
+        cls.server.shutdown()
+        cls.server.server_close()
+        cls.thread.join()
+
+    def setUp(self):
+        self.context = self.browser.new_context(viewport={'width': 1440, 'height': 1000})
+        self.page = self.context.new_page()
+        self.errors = []
+        self.requests = []
+        self.page.on('pageerror', lambda error: self.errors.append(str(error)))
+        self.page.on('request', lambda request: self.requests.append((request.method, request.url)))
+
+    def tearDown(self):
+        self.context.close()
+        self.assertEqual(self.errors, [])
+
+    def open(self, state='ready'):
+        self.page.goto(self.base, wait_until='domcontentloaded')
+        self.page.wait_for_selector('#dataStatus[data-state="' + state + '"]')
+
+    def refresh(self, state='ready', mobile=False):
+        with self.page.expect_response('**/front/data/benchmark.json'):
+            self.page.locator('#refreshMobileBtn' if mobile else '#refreshBtn').click()
+        self.page.wait_for_selector('#dataStatus[data-state="' + state + '"]')
+
+    def test_published_scores_filters_chart_and_export(self):
+        self.open()
+        self.assertEqual(self.page.locator('#kpiModels').inner_text(), str(len(SNAPSHOT['models'])))
+        self.assertEqual(self.page.locator('#datasetFilters input').count(), len(SNAPSHOT['datasets']))
+        self.assertEqual(self.page.locator('[data-tab="submit"], #submitForm').count(), 0)
+        self.assertFalse(any('/api/' in url or method != 'GET' for method, url in self.requests))
+        # Check rendered means against the preserved scoring rules, independently in Python.
+        excluded = {'멀티 모달', '팩트 체크', 'K-Prism (Text, Intact)',
+                    'K-Prism (Text, Perturbed)', 'K-Prism (Image, Intact)', 'K-Prism (Image, Perturbed)'}
+        rows = self.page.locator('#tableBody tr').all()
+        for model in SNAPSHOT['models']:
+            cells = next(row.locator('td').all() for row in rows
+                         if row.locator('td').nth(2).locator('.font-semibold').inner_text() == model['name'])
+            scores = {score['dataset_name']: score['score'] for score in model['scores']}
+            datasets = [d for d in SNAPSHOT['datasets'] if model['is_multimodal'] or d not in excluded]
+            expected = sum(scores[d] for d in datasets) / len(datasets)
+            self.assertAlmostEqual(float(cells[3].inner_text()), expected, delta=0.006)
+        self.page.locator('button[data-fav]').first.click()
+        self.page.wait_for_function("window.Chart && Chart.getChart('scoresChart').data.datasets.length === 2")
+        self.page.locator('#searchInput').fill('qwen3-32b')
+        self.assertEqual(self.page.locator('#kpiModels').inner_text(), '1')
+        with self.page.expect_download() as download:
+            self.page.locator('#exportCsvBtn').click()
+        rows = list(csv.reader(io.StringIO(Path(download.value.path()).read_text())))
+        self.assertEqual(len(rows), 2)
+        self.assertEqual(rows[1][2], 'qwen3-32b')
+        self.page.locator('#searchInput').fill('')
+        self.page.locator('#providerSelect').select_option('OpenAI')
+        self.assertEqual(self.page.locator('#kpiModels').inner_text(), '1')
+        self.page.locator('#searchInput').fill('no-model-matches-this')
+        self.assertIn('No models match', self.page.locator('#tableBody').inner_text())
+
+    def test_preferences_and_repeated_refresh(self):
+        self.open()
+        self.page.locator('#clearAllBtn').click()
+        self.refresh()
+        self.refresh()
+        self.assertEqual(self.page.locator('#datasetFilters input:checked').count(), 0)
+        self.page.reload(wait_until='domcontentloaded')
+        self.page.wait_for_selector('#dataStatus[data-state="ready"]')
+        self.assertEqual(self.page.locator('#datasetFilters input:checked').count(), 0)
+        self.page.locator('#selectAllBtn').click()
+        self.page.locator('#searchInput').fill('qwen3-32b')
+        self.page.reload(wait_until='domcontentloaded')
+        self.page.wait_for_selector('#dataStatus[data-state="ready"]')
+        self.assertEqual(self.page.locator('#searchInput').input_value(), 'qwen3-32b')
+        self.assertEqual(self.page.locator('#kpiModels').inner_text(), '1')
+
+    def test_initial_failure_never_invents_results_and_retry_recovers(self):
+        self.page.route('**/front/data/benchmark.json', lambda route: route.fulfill(status=503, body='Unavailable'))
+        self.open('error')
+        self.assertEqual(self.page.locator('#kpiModels').inner_text(), '—')
+        self.assertEqual(self.page.locator('button[data-fav]').count(), 0)
+        self.assertTrue(self.page.locator('#exportCsvBtn').is_disabled())
+        self.page.unroute('**/front/data/benchmark.json')
+        self.refresh()
+        self.assertEqual(self.page.locator('#kpiModels').inner_text(), str(len(SNAPSHOT['models'])))
+
+    def test_malformed_refresh_preserves_previous_results(self):
+        self.open()
+        before = self.page.locator('#tableBody').inner_text()
+        self.page.route('**/front/data/benchmark.json', lambda route: route.fulfill(
+            content_type='application/json', body='{"models":[{"provider":"Broken"}]}'))
+        self.refresh('error')
+        self.assertEqual(self.page.locator('#tableBody').inner_text(), before)
+        self.assertIn('Previously loaded', self.page.locator('#dataStatus').inner_text())
+
+    def test_mobile_controls(self):
+        self.page.set_viewport_size({'width': 390, 'height': 844})
+        self.open()
+        self.assertTrue(self.page.locator('#refreshMobileBtn').is_visible())
+        self.assertLessEqual(self.page.evaluate('document.documentElement.scrollWidth'), 390)
+        mean = self.page.locator('#tableBody tr').first.locator('td.mean-col').bounding_box()
+        self.assertLessEqual(mean['x'] + mean['width'], 390)
+        before = self.page.locator('html').get_attribute('class') or ''
+        self.page.locator('#themeToggleMobile').click()
+        self.assertNotEqual(self.page.locator('html').get_attribute('class') or '', before)
+        self.refresh(mobile=True)
+        self.assertEqual(self.page.locator('#kpiModels').inner_text(), str(len(SNAPSHOT['models'])))
+
+    def test_blocked_storage_and_unavailable_chart_keep_table_usable(self):
+        self.page.add_init_script("Object.defineProperty(window, 'localStorage', {get() { throw new Error('blocked'); }});")
+        self.page.route('https://cdn.jsdelivr.net/npm/chart.js', lambda route: route.abort())
+        self.open()
+        self.assertTrue(self.page.locator('#chartUnavailable').is_visible())
+        self.page.locator('#searchInput').fill('qwen3-32b')
+        self.assertEqual(self.page.locator('#kpiModels').inner_text(), '1')
+        self.page.locator('#themeToggle').click()
+
+    def test_empty_results_and_untrusted_labels(self):
+        self.page.route('**/front/data/benchmark.json', lambda route: route.fulfill(
+            content_type='application/json', body='{"models":[],"datasets":[]}'))
+        self.open()
+        self.assertEqual(self.page.locator('#kpiModels').inner_text(), '0')
+        self.assertIn('No results have been published', self.page.locator('#tableBody').inner_text())
+        self.page.unroute('**/front/data/benchmark.json')
+        label = '<img src=x onerror="window.injected=true">'
+        payload = {'datasets': [label], 'models': [{
+            'provider': label, 'name': label, 'repo': label, 'is_multimodal': True,
+            'scores': [{'dataset_name': label, 'score': 42, 'metric_type': 'raw'}]}]}
+        self.page.route('**/front/data/benchmark.json', lambda route: route.fulfill(
+            content_type='application/json', body=json.dumps(payload)))
+        self.refresh()
+        self.page.locator('#selectAllBtn').click()
+        self.assertEqual(self.page.locator('#tableBody img, #datasetFilters img').count(), 0)
+        self.assertFalse(self.page.evaluate('Boolean(window.injected)'))
+        self.assertIn(label, self.page.locator('#tableBody').inner_text())
+
+    def test_legacy_array_validation_and_request_timeout(self):
+        self.open()
+        result = self.page.evaluate('''async () => {
+          const {normalizeBenchmark, loadBenchmark} = await import('/front/data.mjs');
+          const legacy = normalizeBenchmark([{provider: 'P', name: 'M', scores: [
+            {dataset_name: 'Additional', score: 0}]}]);
+          let invalid = 0;
+          for (const payload of [null, {}, {models: [{provider:'P', name:'M', scores:[
+            {dataset_name:'A', score:'99'}]}]}, {models:[], datasets:'invalid'}]) {
+            try { normalizeBenchmark(payload); } catch { invalid++; }
+          }
+          window.fetch = (_url, options) => new Promise((_resolve, reject) => {
+            options.signal.addEventListener('abort', () => reject(new DOMException('Aborted', 'AbortError')));
+          });
+          let timedOut = false;
+          try { await loadBenchmark('/never-responds', 20); } catch (error) { timedOut = error.name === 'AbortError'; }
+          return {datasets: legacy.datasets, score: legacy.models[0].scores[0].score, invalid, timedOut};
+        }''')
+        self.assertIn('Additional', result['datasets'])
+        self.assertEqual(result['score'], 0)
+        self.assertEqual(result['invalid'], 4)
+        self.assertTrue(result['timedOut'])
+
+
+if __name__ == '__main__':
+    unittest.main()