fix: use actual text length for HRM input instead of padding to 2048

index.html

@@ -521,16 +521,15 @@
     const bytes = encoder.encode(text);
     const len = Math.min(bytes.length, HRM_MAX_LEN);
 
-    const inputIds = new BigInt64Array(HRM_MAX_LEN);
-    const attentionMask = new BigInt64Array(HRM_MAX_LEN);
+    const inputIds = new BigInt64Array(len);
+    const attentionMask = new BigInt64Array(len);
 
     for (let i = 0; i < len; i++) {
       inputIds[i] = BigInt(bytes[i]);
       attentionMask[i] = 1n;
     }
-    // Remaining positions are already 0 (padded)
 
-    return { inputIds, attentionMask };
+    return { inputIds, attentionMask, seqLen: len };
   }
 
   async function analyzeDistilbert(text) {
@@ -540,11 +539,11 @@
   }
 
   async function analyzeHrm(text) {
-    const { inputIds, attentionMask } = tokenizeBytes(text);
+    const { inputIds, attentionMask, seqLen } = tokenizeBytes(text);
     const ortLib = await getOrt();
 
-    const inputTensor = new ortLib.Tensor('int64', inputIds, [1, HRM_MAX_LEN]);
-    const maskTensor = new ortLib.Tensor('int64', attentionMask, [1, HRM_MAX_LEN]);
+    const inputTensor = new ortLib.Tensor('int64', inputIds, [1, seqLen]);
+    const maskTensor = new ortLib.Tensor('int64', attentionMask, [1, seqLen]);
 
     const results = await hrmSession.run({
       input_ids: inputTensor,