From bde21f6624b23bbb2e24bb470dd989183032c06e Mon Sep 17 00:00:00 2001 From: sora <2075279110@qq.com> Date: Wed, 29 Jul 2026 08:19:04 +0000 Subject: [PATCH] Replace meaningless trend charts with overall rank and win matrix in results page --- webui/results_scan.py | 51 ++++++++++++++++++++++++--- webui/static/results.html | 37 ++++++++++---------- webui/static/results.js | 72 +++++++++++++++++++++++++++++++-------- 3 files changed, 124 insertions(+), 36 deletions(-) diff --git a/webui/results_scan.py b/webui/results_scan.py index 7b7888d..1ed74bb 100644 --- a/webui/results_scan.py +++ b/webui/results_scan.py @@ -9,18 +9,18 @@ from typing import Dict, List, Optional, Tuple # Display order for comparison charts CATEGORY_ORDER = [ - ('math', '数学推理', [ + ('math', '数学', [ 'aime24', 'aime25', 'aime26', 'hmmt26', 'imo_answerbench', 'competition_math', 'gsm8k', ]), ('code', '代码', ['humaneval', 'live_code_bench', 'bigcodebench']), - ('science', '科学 / 高难推理', ['gpqa_diamond', 'super_gpqa', 'hle']), - ('knowledge', '知识与通用能力', [ + ('science', '科学', ['gpqa_diamond', 'super_gpqa', 'hle']), + ('knowledge', '知识', [ 'mmlu', 'mmlu_pro', 'cmmlu', 'bbh', 'arc', 'drop', 'hellaswag', 'winogrande', 'simple_qa', 'trivia_qa', ]), ('long_context', '长文本', ['longbench_v2', 'openai_mrcr']), - ('tool_agent', '工具调用 / 智能体', ['bfcl_v3', 'general_fc', 'tau2_bench']), + ('tool_agent', '智能体', ['bfcl_v3', 'general_fc', 'tau2_bench']), ] BENCHMARK_ALIAS = { @@ -367,6 +367,46 @@ def compare_models( 'rows': rows, }) + # Overall aggregates + overall = [] + for s in series: + vals = [v for v in s['scores'] if v is not None] + overall.append({ + 'folder': s['folder'], + 'label': s['label'], + 'display': s['display'], + 'mean': round(sum(vals) / len(vals), 6) if vals else None, + 'n': len(vals), + }) + overall.sort(key=lambda x: (x['mean'] if x['mean'] is not None else -1), reverse=True) + + # Win count: how many benchmarks each model ranks #1 + win_counts = {s['folder']: 0 for s in series} + for i, b in enumerate(bench_list): + best_score = -1 + best_folders = [] + for s in series: + v = s['scores'][i] + if v is None: + continue + if v > best_score: + best_score = v + best_folders = [s['folder']] + elif v == best_score: + best_folders.append(s['folder']) + for f in best_folders: + win_counts[f] += 1 / len(best_folders) + + # Head-to-head win matrix + folders = [s['folder'] for s in series] + win_matrix = {f: {g: 0 for g in folders} for f in folders} + for i, b in enumerate(bench_list): + scores_here = [(s['folder'], s['scores'][i]) for s in series if s['scores'][i] is not None] + for f, sv in scores_here: + for g, gv in scores_here: + if sv > gv: + win_matrix[f][g] += 1 + return { 'output_dir': overview['output_dir'], 'benchmarks': bench_list, @@ -377,6 +417,9 @@ def compare_models( 'category_labels': category_labels, 'category_series': category_series, 'category_ranking': category_ranking, + 'overall': overall, + 'win_counts': win_counts, + 'win_matrix': win_matrix, 'matrix': { f: {b: overview['matrix'].get(f, {}).get(b) for b in bench_list} for f in selected diff --git a/webui/static/results.html b/webui/static/results.html index 002ea3b..f0613cc 100644 --- a/webui/static/results.html +++ b/webui/static/results.html @@ -67,17 +67,15 @@
-

能力分类走势

- 横轴 = 能力域，纵轴 = 该域平均分 -
-
- +

综合排名

+ 所选 benchmark 平均分 + 冠军次数
+
-

能力分类柱状对比

+

能力分类对比

各模型在同一能力域的平均得分
@@ -87,17 +85,7 @@
-

得分走势

- 横轴 = benchmark，纵轴 = score -
-
- -
-
- -
-
-

柱状对比

+

各 Benchmark 对比

同 benchmark 下各模型得分
@@ -115,6 +103,19 @@
+
+
+

胜负矩阵

+ 行模型在多少 benchmark 上击败列模型 +
+
+ + + +
+
+
+

得分表

@@ -137,6 +138,6 @@
- + diff --git a/webui/static/results.js b/webui/static/results.js index 038fcd6..cc9fc50 100644 --- a/webui/static/results.js +++ b/webui/static/results.js @@ -2,7 +2,7 @@ const state = { overview: null, compare: null, - charts: { line: null, bar: null, radar: null, catLine: null, catBar: null }, + charts: { bar: null, radar: null, catBar: null }, }; const $ = (id) => document.getElementById(id); @@ -197,17 +197,10 @@ const catLabels = compare.category_labels || []; const catSeries = compare.category_series || []; - const catLineCanvas = $('categoryLineChart'); const catBarCanvas = $('categoryBarChart'); if (!catLabels.length || !catSeries.length) { - markEmpty(catLineCanvas, '当前选择下无可用能力分类数据'); markEmpty(catBarCanvas, '当前选择下无可用能力分类数据'); } else { - makeChart('catLine', catLineCanvas, { - type: 'line', - data: { labels: catLabels, datasets: chartDatasets(catSeries, 'line') }, - options: { ...common, scales: axisOptions(0) }, - }); makeChart('catBar', catBarCanvas, { type: 'bar', data: { labels: catLabels, datasets: chartDatasets(catSeries, 'bar') }, @@ -222,14 +215,8 @@ } if (!labels.length) { - markEmpty($('lineChart'), '暂无 benchmark 得分'); markEmpty($('barChart'), '暂无 benchmark 得分'); } else { - makeChart('line', $('lineChart'), { - type: 'line', - data: { labels, datasets: chartDatasets(compare.series, 'line') }, - options: { ...common, scales: axisOptions(45) }, - }); makeChart('bar', $('barChart'), { type: 'bar', data: { labels, datasets: chartDatasets(compare.series, 'bar') }, @@ -306,6 +293,61 @@ }); } + function renderWinMatrix(compare) { + const table = $('winMatrixTable'); + const thead = table.querySelector('thead'); + const tbody = table.querySelector('tbody'); + thead.innerHTML = ''; + tbody.innerHTML = ''; + + const folders = compare.series.map((s) => s.folder); + if (folders.length < 2) { + tbody.innerHTML = '至少选择 2 个模型才能生成胜负矩阵'; + return; + } + const matrix = compare.win_matrix || {}; + + const head = document.createElement('tr'); + head.innerHTML = `模型${folders.map((f) => `${f}`).join('')}`; + thead.appendChild(head); + + folders.forEach((f) => { + const tr = document.createElement('tr'); + const cells = folders.map((g) => { + const val = matrix[f]?.[g] ?? 0; + return `${val}`; + }).join(''); + tr.innerHTML = `${f}${cells}`; + tbody.appendChild(tr); + }); + } + + function renderOverallRank(compare) { + const box = $('overallRank'); + box.innerHTML = ''; + const overall = compare.overall || []; + const wins = compare.win_counts || {}; + if (!overall.length) { + box.innerHTML = '
无数据
'; + return; + } + overall.forEach((item, idx) => { + const card = document.createElement('div'); + card.className = 'rank-card'; + const pct = item.mean != null ? (item.mean * 100).toFixed(2) + '%' : '—'; + const win = wins[item.folder] != null ? wins[item.folder].toFixed(1) : '0'; + card.innerHTML = ` +
+ #${idx + 1} + ${item.folder} + ${pct} +
+ ${item.n} benches · ${win} 项第一 + `; + box.appendChild(card); + }); + } + function renderRanking(compare) { const box = $('rankList'); box.innerHTML = ''; @@ -364,6 +406,8 @@ state.compare = data; renderCharts(data); renderTable(data); + renderWinMatrix(data); + renderOverallRank(data); renderRanking(data); setMsg(`已对比 ${data.series.length} 模型 × ${data.benchmarks.length} benches`, 'ok'); }