maxidl Claude Opus 5.5 commited on
Commit Β·
f72f04c
1
Parent(s): cd4d848
Metric selector: separate strict-match and flexible-extract rows
Browse filesThe *_gsm8k_platinum_cot tasks have exact_match rows for two answer filters.
The chart queried by task and metric only, so each series drew both rows per
step. Metrics with several filters now get one option per filter
(strict-match first, the filter the suite aggregates use), and the queries
filter on metric_filter.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- index.html +29 -7
index.html
CHANGED
|
@@ -552,6 +552,15 @@
|
|
| 552 |
function esc(s) { return s.replace(/'/g, "''"); }
|
| 553 |
function sqlIn(vals) { return vals.map(v => `'${esc(v)}'`).join(', '); }
|
| 554 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 555 |
// ββ Shared helpers ββββββββββββββββββββββββββββββββββββββββββ
|
| 556 |
function populateSelect(el, options, selected) {
|
| 557 |
el.innerHTML = '';
|
|
@@ -1322,10 +1331,23 @@
|
|
| 1322 |
if (!task) return;
|
| 1323 |
|
| 1324 |
const rows = await query(`
|
| 1325 |
-
SELECT DISTINCT metric
|
|
|
|
| 1326 |
`);
|
| 1327 |
-
const
|
| 1328 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1329 |
if (defaults?.chartType) this.el.chartType.value = defaults.chartType;
|
| 1330 |
await this.renderChart();
|
| 1331 |
}
|
|
@@ -1341,7 +1363,7 @@
|
|
| 1341 |
}
|
| 1342 |
const available = await query(`
|
| 1343 |
SELECT DISTINCT model_display_name FROM scores
|
| 1344 |
-
WHERE task = '${esc(task)}' AND
|
| 1345 |
`);
|
| 1346 |
const availableSet = new Set(available.map(r => r.model_display_name));
|
| 1347 |
nameEls.forEach(el => {
|
|
@@ -1381,7 +1403,7 @@
|
|
| 1381 |
is_checkpoint, higher_is_better, step
|
| 1382 |
FROM scores
|
| 1383 |
WHERE task = '${esc(task)}'
|
| 1384 |
-
AND
|
| 1385 |
AND model_display_name IN (${sqlIn(models)})
|
| 1386 |
ORDER BY model_display_name, tokens_trained
|
| 1387 |
`);
|
|
@@ -1403,7 +1425,7 @@
|
|
| 1403 |
try {
|
| 1404 |
const stRows = await query(`
|
| 1405 |
SELECT DISTINCT subtask_tree FROM scores
|
| 1406 |
-
WHERE task = '${esc(task)}' AND
|
| 1407 |
AND subtask_tree IS NOT NULL
|
| 1408 |
`);
|
| 1409 |
if (stRows.length > 0) {
|
|
@@ -1485,7 +1507,7 @@
|
|
| 1485 |
|
| 1486 |
formatChartTitle(task, metric, higherIsBetter) {
|
| 1487 |
const arrow = higherIsBetter === true ? ' \u2191' : higherIsBetter === false ? ' \u2193' : '';
|
| 1488 |
-
return `${task} \u2014 ${metric}${arrow}`;
|
| 1489 |
}
|
| 1490 |
|
| 1491 |
renderSubtaskTree(map, keys, depth = 0) {
|
|
|
|
| 552 |
function esc(s) { return s.replace(/'/g, "''"); }
|
| 553 |
function sqlIn(vals) { return vals.map(v => `'${esc(v)}'`).join(', '); }
|
| 554 |
|
| 555 |
+
// A metric option is "metric", or "metric|filter" when the task has rows for several answer filters
|
| 556 |
+
// (e.g. *_gsm8k_platinum_cot exact_match: strict-match and flexible-extract)
|
| 557 |
+
function splitMetric(v) { const i = v.indexOf('|'); return i < 0 ? [v, null] : [v.slice(0, i), v.slice(i + 1)]; }
|
| 558 |
+
function metricWhere(v) {
|
| 559 |
+
const [m, f] = splitMetric(v);
|
| 560 |
+
return `metric = '${esc(m)}'` + (f ? ` AND COALESCE(metric_filter, 'none') = '${esc(f)}'` : '');
|
| 561 |
+
}
|
| 562 |
+
function metricLabel(v) { const [m, f] = splitMetric(v); return f ? `${m} (${f})` : m; }
|
| 563 |
+
|
| 564 |
// ββ Shared helpers ββββββββββββββββββββββββββββββββββββββββββ
|
| 565 |
function populateSelect(el, options, selected) {
|
| 566 |
el.innerHTML = '';
|
|
|
|
| 1331 |
if (!task) return;
|
| 1332 |
|
| 1333 |
const rows = await query(`
|
| 1334 |
+
SELECT DISTINCT metric, COALESCE(metric_filter, 'none') AS metric_filter
|
| 1335 |
+
FROM scores WHERE task = '${esc(task)}' ORDER BY metric, metric_filter
|
| 1336 |
`);
|
| 1337 |
+
const filters = new Map();
|
| 1338 |
+
rows.forEach(r => filters.set(r.metric, [...(filters.get(r.metric) || []), r.metric_filter]));
|
| 1339 |
+
const options = [];
|
| 1340 |
+
for (const [m, fs] of filters) {
|
| 1341 |
+
if (fs.length < 2) { options.push(m); continue; }
|
| 1342 |
+
// strict-match first: it is the filter the suite aggregates use
|
| 1343 |
+
fs.sort((a, b) => (b === 'strict-match') - (a === 'strict-match') || a.localeCompare(b));
|
| 1344 |
+
fs.forEach(f => options.push({ value: `${m}|${f}`, label: `${m} (${f})` }));
|
| 1345 |
+
}
|
| 1346 |
+
let prev = defaults?.metric || this.el.metric.value;
|
| 1347 |
+
if (prev && !prev.includes('|') && (filters.get(prev) || []).length > 1) {
|
| 1348 |
+
prev = options.find(o => typeof o === 'object' && o.value.startsWith(prev + '|')).value;
|
| 1349 |
+
}
|
| 1350 |
+
populateSelect(this.el.metric, options, prev);
|
| 1351 |
if (defaults?.chartType) this.el.chartType.value = defaults.chartType;
|
| 1352 |
await this.renderChart();
|
| 1353 |
}
|
|
|
|
| 1363 |
}
|
| 1364 |
const available = await query(`
|
| 1365 |
SELECT DISTINCT model_display_name FROM scores
|
| 1366 |
+
WHERE task = '${esc(task)}' AND ${metricWhere(metric)}
|
| 1367 |
`);
|
| 1368 |
const availableSet = new Set(available.map(r => r.model_display_name));
|
| 1369 |
nameEls.forEach(el => {
|
|
|
|
| 1403 |
is_checkpoint, higher_is_better, step
|
| 1404 |
FROM scores
|
| 1405 |
WHERE task = '${esc(task)}'
|
| 1406 |
+
AND ${metricWhere(metric)}
|
| 1407 |
AND model_display_name IN (${sqlIn(models)})
|
| 1408 |
ORDER BY model_display_name, tokens_trained
|
| 1409 |
`);
|
|
|
|
| 1425 |
try {
|
| 1426 |
const stRows = await query(`
|
| 1427 |
SELECT DISTINCT subtask_tree FROM scores
|
| 1428 |
+
WHERE task = '${esc(task)}' AND ${metricWhere(metric)}
|
| 1429 |
AND subtask_tree IS NOT NULL
|
| 1430 |
`);
|
| 1431 |
if (stRows.length > 0) {
|
|
|
|
| 1507 |
|
| 1508 |
formatChartTitle(task, metric, higherIsBetter) {
|
| 1509 |
const arrow = higherIsBetter === true ? ' \u2191' : higherIsBetter === false ? ' \u2193' : '';
|
| 1510 |
+
return `${task} \u2014 ${metricLabel(metric)}${arrow}`;
|
| 1511 |
}
|
| 1512 |
|
| 1513 |
renderSubtaskTree(map, keys, depth = 0) {
|