+
${escapeHtml(job.id)}
${escapeHtml(job.status)}
-
${escapeHtml(model)} · ${thinking}
+
${escapeHtml(model)} · ${mode} · ${escapeHtml(scope)}
`;
btn.addEventListener('click', () => followJob(job.id, true));
- if (job.id === activeId) btn.style.borderColor = 'rgba(214,162,74,0.75)';
jobList.appendChild(btn);
});
}
@@ -355,6 +395,19 @@
return data;
}
+ function tryParseProgress(text) {
+ // Look for patterns like "Evaluating[mmlu_pro] 12%| 1450/12032"
+ const m = text.match(/\[eval\]\s*(\d+)%\|\s*(\d+)\/(\d+)/i);
+ if (m) {
+ return { pct: Number(m[1]), current: Number(m[2]), total: Number(m[3]) };
+ }
+ const m2 = text.match(/Evaluating\[(\w+)\]\s+(\d+)%\|\s*(\d+)\/(\d+)/);
+ if (m2) {
+ return { pct: Number(m2[2]), current: Number(m2[3]), total: Number(m2[4]) };
+ }
+ return null;
+ }
+
async function pullLogs(reset = false) {
if (!state.activeJobId) return;
if (reset) {
@@ -366,12 +419,22 @@
logView.textContent += data.content;
state.logOffset = data.next_offset;
if ($('autoScroll').checked) logView.scrollTop = logView.scrollHeight;
+
+ const prog = tryParseProgress(logView.textContent);
+ if (prog) {
+ $('progressBar').style.width = `${prog.pct}%`;
+ $('progressText').textContent = `${prog.pct}% · ${prog.current}/${prog.total}`;
+ }
}
if (data.status) {
$('activeStatus').className = `badge ${statusClass(data.status)}`;
$('activeStatus').textContent = data.status;
$('activeCard').className = `active-card ${statusClass(data.status)}`;
$('stopBtn').disabled = data.status !== 'running';
+ if (data.done) {
+ $('progressBar').style.width = '100%';
+ $('progressText').textContent = data.status === 'completed' ? '已完成' : `结束: ${data.status}`;
+ }
}
if (data.done) stopPolling();
}
@@ -403,114 +466,26 @@
else stopPolling();
}
- async function init() {
- try {
- await api('/api/health');
- $('healthDot').className = 'dot ok';
- $('healthText').textContent = 'server online';
- } catch (e) {
- $('healthDot').className = 'dot bad';
- $('healthText').textContent = 'server offline';
- setMsg(e.message, 'error');
- return;
- }
-
- state.meta = await api('/api/meta');
- fillDefaults(state.meta);
- renderSuites();
- renderCustomSuites();
- renderBenchmarks();
- renderCustomPicker();
- setMode('suite');
-
- const jobs = await refreshJobs();
- if (jobs.active_job_id) {
- await followJob(jobs.active_job_id, true);
- } else if (jobs.jobs?.[0]) {
- await followJob(jobs.jobs[0].id, true);
- }
- }
-
- document.querySelectorAll('.mode-tabs .tab').forEach((el) => {
- el.addEventListener('click', () => setMode(el.dataset.mode));
- });
- $('selectAllBtn').addEventListener('click', () => {
- document.querySelectorAll('#benchmarkList input').forEach((el) => { el.checked = true; });
- updateSelectedCount();
- });
- $('clearAllBtn').addEventListener('click', () => {
- document.querySelectorAll('#benchmarkList input').forEach((el) => { el.checked = false; });
- updateSelectedCount();
- });
- $('refreshJobsBtn').addEventListener('click', () => refreshJobs().catch((e) => setMsg(e.message, 'error')));
-
- $('saveCustomSuiteBtn').addEventListener('click', async () => {
- const name = $('customSuiteName').value.trim();
- const datasets = [...document.querySelectorAll('#customPickList input:checked')].map((el) => el.value);
- if (!name) {
- setMsg('请填写组合名称', 'error');
- return;
- }
- if (!datasets.length) {
- setMsg('请至少勾选一个 benchmark', 'error');
- return;
- }
- try {
- await saveCustomSuite(name, datasets);
- } catch (err) {
- setMsg(err.message, 'error');
- }
- });
-
- $('loadCheckedToCustomBtn').addEventListener('click', () => {
- // no-op helper text: already editing in place; just focus name
- $('customSuiteName').focus();
- setMsg('请在下方勾选数据集并填写组合名称后保存', 'ok');
- });
-
- $('saveFromDatasetsBtn').addEventListener('click', async () => {
- const datasets = [...document.querySelectorAll('#benchmarkList input:checked')].map((el) => el.value);
- if (!datasets.length) {
- setMsg('请先勾选 benchmark', 'error');
- return;
- }
- const name = prompt('输入要永久保存的组合名称：');
- if (!name || !name.trim()) return;
- try {
- // sync into custom picker too
- const set = new Set(datasets);
- document.querySelectorAll('#customPickList input').forEach((el) => {
- el.checked = set.has(el.value);
- });
- $('customSuiteName').value = name.trim();
- updateCustomPickCount();
- await saveCustomSuite(name.trim(), datasets);
- setMode('suite');
- } catch (err) {
- setMsg(err.message, 'error');
- }
- });
-
- form.addEventListener('submit', async (e) => {
- e.preventDefault();
+ async function createJob() {
setMsg('');
- const requiredPaths = [
+ const required = [
['dataset_dir', 'Dataset Dir'],
['output_dir', 'Output Dir'],
['config', 'Config YAML'],
['tokenizer_path', 'Tokenizer Path'],
];
- for (const [id, label] of requiredPaths) {
+ for (const [id, label] of required) {
if (!$(id).value.trim()) {
setMsg(`请填写必填路径: ${label}`, 'error');
+ setTab('config');
$(id).focus();
return;
}
}
-
- const payload = collectPayload();
+ const payload = buildCommandPayload();
if (payload.selection_mode === 'datasets' && !payload.datasets.length) {
- setMsg('请至少选择一个 benchmark / 自定义组合', 'error');
+ setMsg('请至少选择一个 benchmark', 'error');
+ setTab('scope');
return;
}
$('launchBtn').disabled = true;
@@ -531,8 +506,139 @@
} finally {
$('launchBtn').disabled = false;
}
+ }
+
+ // ---- Presets (localStorage) ----
+ const PRESET_KEY = 'evalstone_path_presets';
+ function getPresets() {
+ try {
+ return JSON.parse(localStorage.getItem(PRESET_KEY) || '[]');
+ } catch {
+ return [];
+ }
+ }
+ function savePreset() {
+ const name = prompt('预设名称');
+ if (!name) return;
+ const presets = getPresets();
+ presets.push({
+ name,
+ dataset_dir: $('dataset_dir').value,
+ output_dir: $('output_dir').value,
+ config: $('config').value,
+ tokenizer_path: $('tokenizer_path').value,
+ });
+ localStorage.setItem(PRESET_KEY, JSON.stringify(presets));
+ setMsg('预设已保存', 'ok');
+ }
+ function loadPreset() {
+ const presets = getPresets();
+ if (!presets.length) {
+ setMsg('没有保存的预设', 'error');
+ return;
+ }
+ const list = presets.map((p, i) => `${i + 1}. ${p.name}`).join('\n');
+ const idx = prompt(`选择预设编号:\n${list}`);
+ const n = Number(idx);
+ if (!n || n < 1 || n > presets.length) return;
+ const p = presets[n - 1];
+ $('dataset_dir').value = p.dataset_dir || '';
+ $('output_dir').value = p.output_dir || '';
+ $('config').value = p.config || '';
+ $('tokenizer_path').value = p.tokenizer_path || '';
+ setMsg(`已加载预设: ${p.name}`, 'ok');
+ }
+ function applyDefaultPreset() {
+ const d = state.meta.defaults;
+ $('dataset_dir').value = d.dataset_dir || '';
+ $('output_dir').value = d.output_dir || '';
+ $('config').value = d.config || '';
+ $('tokenizer_path').value = d.tokenizer_path || '';
+ }
+
+ // ---- Init ----
+ async function init() {
+ try {
+ await api('/api/health');
+ $('healthDot').className = 'dot ok';
+ $('healthText').textContent = 'online';
+ } catch (e) {
+ $('healthDot').className = 'dot bad';
+ $('healthText').textContent = 'offline';
+ setMsg(e.message, 'error');
+ return;
+ }
+
+ state.meta = await api('/api/meta');
+ fillDefaults(state.meta);
+ renderSuites();
+ renderCustomSuites();
+ renderBenchmarks();
+ renderCategoryChips();
+ setMode('suite');
+ setTab('scope');
+ updateEstimate();
+
+ const jobs = await refreshJobs();
+ if (jobs.active_job_id) {
+ await followJob(jobs.active_job_id, true);
+ } else if (jobs.jobs?.[0]) {
+ await followJob(jobs.jobs[0].id, true);
+ }
+ }
+
+ // ---- Events ----
+ document.querySelectorAll('.tabs .tab').forEach((el) => {
+ el.addEventListener('click', () => setTab(el.dataset.tab));
+ });
+ document.querySelectorAll('.mode-btn').forEach((el) => {
+ el.addEventListener('click', () => setMode(el.dataset.mode));
});
+ $('benchSearch').addEventListener('input', (e) => {
+ state.searchQuery = e.target.value;
+ applyFilters();
+ });
+ $('selectAllBtn').addEventListener('click', () => {
+ document.querySelectorAll('#benchmarkList input:not(.hidden-item input)').forEach((el) => { el.checked = true; });
+ updateSelectedCount();
+ });
+ $('clearAllBtn').addEventListener('click', () => {
+ document.querySelectorAll('#benchmarkList input').forEach((el) => { el.checked = false; });
+ updateSelectedCount();
+ });
+ $('saveCustomBtn').addEventListener('click', async () => {
+ const datasets = [...document.querySelectorAll('#benchmarkList input:checked')].map((el) => el.value);
+ if (!datasets.length) {
+ setMsg('请先勾选 benchmark', 'error');
+ return;
+ }
+ const name = prompt('输入自定义组合名称：');
+ if (!name || !name.trim()) return;
+ try {
+ const data = await api('/api/custom-suites', {
+ method: 'POST',
+ body: JSON.stringify({ name: name.trim(), datasets }),
+ });
+ state.customSuites = data.suites || {};
+ renderCustomSuites();
+ state.suiteKind = 'custom';
+ state.suite = name.trim();
+ renderSuites();
+ setMode('suite');
+ setMsg(`已保存组合: ${name.trim()}`, 'ok');
+ } catch (err) {
+ setMsg(err.message, 'error');
+ }
+ });
+
+ $('refreshJobsBtn').addEventListener('click', () => refreshJobs().catch((e) => setMsg(e.message, 'error')));
+ $('launchBtn').addEventListener('click', createJob);
+ $('copyCmdBtn').addEventListener('click', () => {
+ const text = $('activeCmd').textContent;
+ if (!text || text === '尚未启动任务') return;
+ navigator.clipboard.writeText(text).then(() => setMsg('命令已复制', 'ok')).catch(() => setMsg('复制失败', 'error'));
+ });
$('stopBtn').addEventListener('click', async () => {
if (!state.activeJobId) return;
if (!confirm(`确认停止任务 ${state.activeJobId}？`)) return;
@@ -547,5 +653,9 @@
}
});
+ $('presetDefaultBtn').addEventListener('click', applyDefaultPreset);
+ $('presetCurrentBtn').addEventListener('click', savePreset);
+ $('presetLoadBtn').addEventListener('click', loadPreset);
+
init();
})();
diff --git a/webui/static/index.html b/webui/static/index.html
index f3651db..ac6250d 100644
--- a/webui/static/index.html
+++ b/webui/static/index.html
@@ -6,7 +6,7 @@
EvalStone Launch
-
+
@@ -16,7 +16,7 @@
EvalStone Launch
-
本地评测启动台 · 基于 bash/run.py
+
基于 bash/run.py 的评测启动台