Fix Sentence Transformers Hub lists and show retrieval validation details

app.js

@@ -97,6 +97,8 @@ function hubList(library, task, name, view, count) {
   const sort = HUB_SORTS[view];
   if (!hubLists.has(key)) {
     const query = new URLSearchParams({ library, sort, direction: "-1", limit: PAGE });
+    // The model API uses filter for library tags; library alone is ignored.
+    if (library === "sentence-transformers") query.set("filter", library);
     if (task) query.set("pipeline_tag", task);
     if (name) query.set("search", name);
     for (const field of [sort, "safetensors", "library_name"]) query.append("expand[]", field);
@@ -107,9 +109,10 @@ function hubList(library, task, name, view, count) {
     while (list.models.length < count && list.next) {
       const response = await fetchOk(list.next);
       list.models.push(...(await response.json())
-        // The Hub's library filter matches a tag that models of other libraries
-        // carry too, such as diffusers' FLUX or sentence-transformers' MiniLM.
-        .filter((model) => model.library_name === library)
+        // Sentence Transformers supports tagged models even when their primary
+        // library is transformers (for example, EmbeddingGemma 2). Match its Hub
+        // listing; retain the primary-library restriction for the other tabs.
+        .filter((model) => library === "sentence-transformers" || model.library_name === library)
         .map((model) => ({ id: model.id, score: model[sort] ?? 0, params: model.safetensors?.total })));
       list.next = response.headers.get("Link")?.match(/<([^>]+)>;\s*rel="next"/)?.[1];
     }
@@ -215,6 +218,11 @@ async function renderPanel() {
         select(active, "version", versionsOf(s.sdk)),
         nameField),
     ...(selected ? [h("p", { class: "pair" }, `neuronx-cc ${latest.sdk} · ${active} ${latest.version}`)] : [tasks]),
+    ...(active === "sentence-transformers" ? [h("p", { class: "intro" },
+      "Text retrieval with SentenceTransformer, CrossEncoder, and MultiVectorEncoder. "
+      + "The query asks which planet is the Red Planet; documents are Venus, Mars, Jupiter, and Saturn. "
+      + "PASS requires finite scores, Mars ranked first, and CPU score agreement (atol=rtol=0.001). "
+      + "Open a run for scores, precision, and configuration details. SparseEncoder is not covered.")] : []),
     results);
 
   let runs, rank, exclusions, taskList;
@@ -343,13 +351,20 @@ function showRun(run) {
     h("p", {}, `${badge(run.result)} · ${run.neuron_cores} NeuronCore${run.neuron_cores > 1 ? "s" : ""}`
       + ` · ${run.runner} · ${run.library} ${run.library_version}`
       + ` · tested ${day(run.tested_at)}`),
+    ...(run.library === "sentence-transformers" ? [
+      h("h3", {}, "Retrieval validation"),
+      h("p", {}, run.notes ?? ""),
+      h("p", {}, `CPU parity: ${run.cpu_match ?? "not recorded"}`),
+      h("p", {}, `Transformers ${run.transformers}; text-only, one logical Neuron device. `
+        + "This checks one retrieval example, not general model quality or throughput."),
+    ] : []),
     h("h3", {}, "Run it"),
     h("p", {}, `On a ${TARGETS[target]?.name ?? target} (${target}) instance with this Neuron stack,`
       + " for example in the AWS Neuron Deep Learning Container:"),
     h("dl", {}, stack.map(([name, version]) => [h("dt", {}, name), h("dd", {}, h("code", {}, version))])),
     h("ol", {},
       pip && h("li", {}, "Install the packages the script adds:", h("pre", {}, pip)),
-      h("li", {}, "Optionally, skip compilation with the graphs this compiler already built:",
+      run.library !== "sentence-transformers" && h("li", {}, "Optionally, skip compilation with the graphs this compiler already built:",
         h("pre", {}, `hf buckets sync ${NEFF_CACHE}/${target}/neuronxcc-${run.neuronx_cc} /tmp/neff_cache`)),
       h("li", {}, "Download ", h("a", { href: scriptUrl, download: "run.py" }, "run.py"), " and run it",
         run.neuron_cores > 1 ? ". It launches its own torch.distributed ranks:" : ":",