banner: BLiMP przeliczony po poprawce harnessu (67000 par), liczby Harta

index.html

@@ -42,10 +42,10 @@
   <div class="sub">Tiny-LM efficiency track (Glint leaderboard). Auto-refreshes every 20s &middot; <span id="ts">&hellip;</span></div>
   <div class="board">
     <span class="tag">ON THE BOARD</span> (Glint Tiny-ML, org <b>Fabryka AI</b>):
-    <b>GoLLeM-v5 64M</b> &mdash; eff 75.86 (BLiMP 75.99 / ARC-E 47.94 / byte_ppl 2.372, maintainer re-benchmark, PR #78) &middot;
-    <b>32M</b> &mdash; eff 75.51 (73.77 / 44.44 / 2.5386) &middot;
-    <b>16M</b> &mdash; eff 74.49 (70.53 / 40.91 / 2.6746).<br>
-    <span class="tag">GOAL: #1 EFF</span> &mdash; JugnuLM-110M-R2+ holds eff 80.20. At 64M we need BLiMP+ARC &gt; 133.5 with byte_ppl ~2.0 (we have 123.9): a ~9.5-point gap, to be closed by data, not size.<br>
+    <b>GoLLeM-v5 64M</b> &mdash; eff 75.81 (BLiMP 75.83 / ARC-E 47.94 / byte_ppl 2.372) &middot;
+    <b>32M</b> &mdash; eff 75.41 (73.48 / 44.44 / 2.5386) &middot;
+    <b>16M</b> &mdash; eff 74.33 (70.08 / 40.91 / 2.6746). BLiMP recomputed on the full 67k-pair set after a fix in our eval harness (it counted 6 phenomena twice); board entries will be corrected.<br>
+    <span class="tag">GOAL: #1 EFF</span> &mdash; JugnuLM-110M-R2+ holds eff 80.20. At 64M we need BLiMP+ARC &gt; 133.5 with byte_ppl ~2.0 (we have 123.8): a ~9.7-point gap, to be closed by data, not size.<br>
     <span class="tag">NOW TRAINING</span> &mdash; <b>data A/B from one shared 64M checkpoint (step 320k)</b>, same seed and recipe: control (ARC-MIX) vs fork-B (ARC-MIX + educational). We found the first fork-B blend confounded (non-shuffled source sampling), so its result is reported as blend-vs-control only; a clean rebuild runs next. Plus a <b>128M</b> scale reference: scale lifts quality, but the size multiplier cancels most of the gain on eff.
     &middot; board: <a href="https://huggingface.co/spaces/Glint-Research/Tiny-ML-Leaderboard">Glint Tiny-ML</a>
   </div>