Adds TinyMoE-100m-2x8-retrained and correcting the results for TinyMoE-100m-2x8

`TinyMoE-100m-2x8-retrained` is trained on a mix of Cosmopedia-v2 and
FineWeb-Edu-dedup (60/40) from the smollm-corpus. Also re-evaluated
`TinyMoE-100m-2x8` with corrected metrics.

For context, I initially misread the leaderboard's terms and used plain
`acc` and bits-per-byte. Both entries now correctly use `acc_norm` for
ARC-Easy and byte perplexity for WikiText-2, so the original model's
scores are lower than what I first reported elsewhere.

Both models have been evaluated using `lm-eval`.

Hope this is fine — let me know if anything else needs adjusting. I do not want to fraudate the result of the model.

index.html

@@ -1254,14 +1254,27 @@ const models = [
     org: "flamef0x",
     params: "99.8M",
     blimp: 61.13,
-    arc: 26.68,
-    wiki: 1.955,
-    tokens: "~625M", // fill in training token count — I don't have this
+    arc: 25.88,
+    wiki: 3.878,
+    tokens: "~625M",
     releaseDate: "2026-06-15",
     links: {
       card: "https://huggingface.co/FlameF0X/TinyMoE-100m-2x8"
     }
   },
+  {
+    name: "TinyMoE-100m-2x8-retrained",
+    org: "flamef0x",
+    params: "99.8M",
+    blimp: 66.01,
+    arc: 33.88,
+    wiki: 2.879,
+    tokens: "—",
+    releaseDate: "2026-07-05",
+    links: {
+      card: "https://huggingface.co/FlameF0X/TinyMoE-100m-2x8-retrained"
+    }
+  },
   {
     name: "SRLM-1M",
     org: "martico2432",