Spaces:
Running
Running
Clean up Source column and per-row dataset label noise
Browse files* Source column: prefer source_metadata.source_name /
source_organization_name (human-readable: 'kaggle', 'Anthropic Eval
Run') over source_data.source_type, which can be a file format like
'Parquet' that's not useful to a reader. Fall back to source_type
only when nothing better is set.
* Per-row dataset label under the score bar: hide it when it equals
the column header (the eval name) and there's no setup label, since
it's pure noise — Global MMLU was rendering 'Global MMLU Lite' under
every bar in a column already titled 'Global MMLU Lite'.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
- components/eval-detail.tsx +22 -11
components/eval-detail.tsx
CHANGED
|
@@ -1149,9 +1149,17 @@ export function EvalDetail({
|
|
| 1149 |
? "THIRD-PARTY"
|
| 1150 |
: "—"
|
| 1151 |
const isThirdParty = evaluatorRel === "third_party"
|
| 1152 |
-
|
| 1153 |
-
|
| 1154 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1155 |
const familyLabel = modelResult.model_info.architecture
|
| 1156 |
?? modelResult.model_info.parameter_count
|
| 1157 |
?? null
|
|
@@ -1261,14 +1269,17 @@ export function EvalDetail({
|
|
| 1261 |
{setupLabel}
|
| 1262 |
</div>
|
| 1263 |
)}
|
| 1264 |
-
{!setupLabel &&
|
| 1265 |
-
|
| 1266 |
-
|
| 1267 |
-
|
| 1268 |
-
|
| 1269 |
-
|
| 1270 |
-
|
| 1271 |
-
|
|
|
|
|
|
|
|
|
|
| 1272 |
</div>
|
| 1273 |
</td>
|
| 1274 |
|
|
|
|
| 1149 |
? "THIRD-PARTY"
|
| 1150 |
: "—"
|
| 1151 |
const isThirdParty = evaluatorRel === "third_party"
|
| 1152 |
+
// Prefer the human-readable source name (e.g. "kaggle",
|
| 1153 |
+
// "Anthropic Eval Run") over `source_data.source_type`,
|
| 1154 |
+
// which can be a file format like "Parquet" that's
|
| 1155 |
+
// useless to a reader. Fall back to source_type only
|
| 1156 |
+
// when nothing else is set.
|
| 1157 |
+
const sourceTypeLabel =
|
| 1158 |
+
modelResult.source_metadata.source_name?.trim() ||
|
| 1159 |
+
modelResult.source_metadata.source_organization_name?.trim() ||
|
| 1160 |
+
(!Array.isArray(modelResult.source_data) && modelResult.source_data.source_type) ||
|
| 1161 |
+
modelResult.source_metadata.source_type ||
|
| 1162 |
+
""
|
| 1163 |
const familyLabel = modelResult.model_info.architecture
|
| 1164 |
?? modelResult.model_info.parameter_count
|
| 1165 |
?? null
|
|
|
|
| 1269 |
{setupLabel}
|
| 1270 |
</div>
|
| 1271 |
)}
|
| 1272 |
+
{!setupLabel &&
|
| 1273 |
+
datasetName &&
|
| 1274 |
+
!isResearchView &&
|
| 1275 |
+
datasetName !== lb.evaluation_name && (
|
| 1276 |
+
<div
|
| 1277 |
+
className="mt-1 font-mono truncate"
|
| 1278 |
+
style={{ fontSize: 10, color: "var(--fg-subtle)" }}
|
| 1279 |
+
>
|
| 1280 |
+
{datasetName}
|
| 1281 |
+
</div>
|
| 1282 |
+
)}
|
| 1283 |
</div>
|
| 1284 |
</td>
|
| 1285 |
|