Upload index.html

index.html

@@ -65,7 +65,7 @@ tr.highlight{background:var(--soft)}
 .version-card h4{margin:8px 0;font-size:21px}.version-card p{margin:0;color:var(--muted);font-size:13px}
 
 .update-log{display:grid;gap:10px;margin-top:30px;max-width:940px}
-.update-item{display:grid;grid-template-columns:150px 1fr;gap:18px;border:1px solid var(--line);background:var(--panel);padding:15px 17px}
+.update-item{display:grid;grid-template-columns:150px 1fr;gap:18px;border:1px solid var(--line);background:var(--panel);padding:15px 17px;cursor:pointer;transition:.18s}.update-item:hover{border-color:rgba(225,29,46,.5);background:var(--panel2);transform:translateY(-1px)}
 .update-date{font:800 11px ui-monospace,monospace;color:var(--accent);letter-spacing:.03em;text-transform:uppercase}
 .update-copy strong{display:block;font-size:14px;margin-bottom:2px}
 .update-copy span{display:block;color:var(--muted);font-size:12px}
@@ -205,7 +205,22 @@ footer{border-top:1px solid var(--line);padding:36px 0;color:var(--muted);font-s
   <div class="wrap detail-hero">
     <div class="eyebrow">Diffusion-style language-model research</div><h2>Eris</h2>
     <p>A non-autoregressive model combining a full-resolution token stream with a compressed global latent stream and parallel refinement.</p>
-    <div class="stats"><div class="stat"><strong>~43M</strong><span>foundation prototype</span></div><div class="stat"><strong>~2.8B</strong><span>tokens at reported eval</span></div><div class="stat"><strong>45.7%</strong><span>top-1 recovery</span></div><div class="stat"><strong>97.2%</strong><span>edit-detection F1</span></div></div><div class="update-log"><div class="update-item"><div class="update-date">18 Aug 2026</div><div class="update-copy"><strong>Architecture and inference tests</strong><span>Synthetic edit benchmarks established the first strong batching-dependent parallel-refinement throughput signal.</span></div></div><div class="update-item"><div class="update-date">19 Aug 2026</div><div class="update-copy"><strong>First pretrained model and first fine-tuning results</strong><span>~43M foundation prototype reached ~2.8B training tokens; first 500-pair real-world correction fine-tune confirmed downstream transfer.</span></div></div></div>
+    <div class="stats"><div class="stat"><strong>~43M</strong><span>foundation prototype</span></div><div class="stat"><strong>~2.8B</strong><span>tokens at reported eval</span></div><div class="stat"><strong>45.7%</strong><span>top-1 recovery</span></div><div class="stat"><strong>97.2%</strong><span>edit-detection F1</span></div></div><div class="update-log">
+      <div class="update-item" onclick="scrollInsideModel('eris-update-2026-08-18')">
+        <div class="update-date">18 Aug 2026</div>
+        <div class="update-copy">
+          <strong>Architecture and inference tests</strong>
+          <span>Synthetic edit benchmarks established the first strong batching-dependent parallel-refinement throughput signal. Click to open the full test note.</span>
+        </div>
+      </div>
+      <div class="update-item" onclick="scrollInsideModel('eris-update-2026-08-19')">
+        <div class="update-date">19 Aug 2026</div>
+        <div class="update-copy">
+          <strong>First pretrained model and first fine-tuning results</strong>
+          <span>~43M foundation prototype, ~2.8B training tokens, C4 ablation and the first 500-pair real-world correction fine-tune. Click for the complete update.</span>
+        </div>
+      </div>
+    </div>
   </div>
 
   <section><div class="wrap">
@@ -216,10 +231,34 @@ footer{border-top:1px solid var(--line);padding:36px 0;color:var(--muted);font-s
     </div>
   </div></section>
 
-  <section><div class="wrap">
-    <div class="eyebrow">Training results</div><h3 class="section-title">Stable large-scale optimization with measurable reconstruction ability.</h3>
+  <section id="eris-update-2026-08-19"><div class="wrap">
+    <div class="eyebrow">19 Aug 2026 · First pretrained model and first fine-tuning results</div>
+    <h3 class="section-title">Eris moved from architecture benchmarks into its first pretrained foundation-model experiment.</h3>
+
+    <div class="prose">
+      <p>
+        The current prototype has approximately <strong>43M parameters</strong>, uses a
+        <strong>~32k Bielik/APT4 tokenizer</strong>, and combines a full-resolution token
+        stream with a <strong>C4-compressed global latent pathway</strong>.
+      </p>
+      <p>
+        At approximately <strong>2.8B training tokens</strong>, the first held-out evaluation
+        measured <strong>reconstruction CE ≈ 3.10</strong>, <strong>45.7% top-1</strong> and
+        <strong>64.1% top-5 token recovery</strong>, with auxiliary edit-detection
+        <strong>F1 ≈ 97.2%</strong>. Training throughput at this stage was roughly
+        <strong>80k tokens/s on an RTX 5060 Ti</strong>.
+      </p>
+    </div>
+
+    <div class="stats">
+      <div class="stat"><strong>~43M</strong><span>foundation prototype</span></div>
+      <div class="stat"><strong>~2.8B</strong><span>pretraining tokens</span></div>
+      <div class="stat"><strong>~80k tok/s</strong><span>RTX 5060 Ti training throughput</span></div>
+      <div class="stat"><strong>~32k</strong><span>Bielik/APT4 tokenizer</span></div>
+    </div>
+
     <div class="table-wrap"><table>
-      <thead><tr><th>Held-out metric at ~2.8B tokens</th><th>Result</th></tr></thead>
+      <thead><tr><th>Pretraining evaluation</th><th>Result</th></tr></thead>
       <tbody>
         <tr class="highlight"><td>Reconstruction CE</td><td>≈ 3.10</td></tr>
         <tr><td>Top-1 token recovery</td><td>45.7%</td></tr>
@@ -227,36 +266,73 @@ footer{border-top:1px solid var(--line);padding:36px 0;color:var(--muted);font-s
         <tr><td>Auxiliary edit-detection F1</td><td>≈ 97.2%</td></tr>
       </tbody>
     </table></div>
-  </div></section>
 
-  <section><div class="wrap">
-    <div class="eyebrow">C4 ablation</div><h3 class="section-title">The compressed global pathway is contributing useful information.</h3>
+    <div class="callout">
+      <strong>C4 global-context ablation</strong>
+      <p>
+        Removing global context degraded reconstruction CE from <strong>≈3.10 to ≈3.21</strong>
+        and top-5 accuracy from <strong>64.1% to 63.0%</strong>. The compressed global branch is
+        therefore contributing useful information rather than behaving like an inactive pathway.
+        Its current compute cost is substantial, however, making cheaper global-context injection
+        one of the main architectural targets.
+      </p>
+    </div>
+
     <div class="table-wrap"><table>
-      <thead><tr><th>Configuration</th><th>Reconstruction CE</th><th>Top-5</th></tr></thead>
+      <thead><tr><th>C4 ablation</th><th>Reconstruction CE</th><th>Top-5 recovery</th></tr></thead>
       <tbody>
-        <tr class="highlight"><td>Full Eris</td><td>≈ 3.10</td><td>64.1%</td></tr>
+        <tr class="highlight"><td>Full Eris prototype</td><td>≈ 3.10</td><td>64.1%</td></tr>
         <tr><td>Global context removed</td><td>≈ 3.21</td><td>63.0%</td></tr>
       </tbody>
     </table></div>
-    <p class="note">The global path helps reconstruction, but its current compute cost is still substantial.</p>
-  </div></section>
 
-  <section><div class="wrap">
-    <div class="eyebrow">Inference</div><h3 class="section-title">Early parallel-refinement throughput is the most striking architectural signal so far.</h3>
-    <div class="prose"><p>At batch size 8, one-pass synthetic edit workloads measured roughly <strong>15×–168×</strong> higher document throughput than similarly sized dense autoregressive baselines, depending on sequence length and workload.</p><p>In tested 1024-token cases, one-pass Eris reached roughly <strong>960–1,122 documents/s</strong> with batch latency around <strong>7.1–8.3 ms</strong>.</p></div>
-    <div class="callout"><strong>Preliminary architecture measurements</strong><p>These results demonstrate stable training, useful global-context learning and promising parallel inference characteristics. They are not yet a finished-model quality comparison against established pretrained language models.</p></div>
-  </div></section>
+    <h3 class="section-title" style="margin-top:44px">First downstream fine-tuning experiment.</h3>
+    <div class="prose">
+      <p>
+        The first downstream test used only <strong>500 manually corrected real-world text pairs</strong>,
+        with <strong>no synthetic corruption</strong>. On a held-out set, Eris achieved roughly
+        <strong>46% exact correction</strong>, improved <strong>69% of corrupted examples</strong>,
+        reduced total edit distance by <strong>49%</strong>, and worsened approximately
+        <strong>6%</strong> of examples.
+      </p>
+      <p>
+        This is a small transfer experiment rather than a finished downstream benchmark, but it
+        suggests that the pretrained representations can transfer to a new task even with very
+        limited supervised data.
+      </p>
+    </div>
 
-  <section><div class="wrap">
-    <div class="eyebrow">19 Aug 2026 · First downstream fine-tune</div><h3 class="section-title">Useful transfer from only 500 manually corrected real-world pairs.</h3>
-    <div class="prose"><p>The first downstream experiment used only <strong>500 manually corrected real-world text pairs</strong>, with no synthetic corruption. This was intended as a small transfer probe rather than a final task-specific model.</p></div>
-    <div class="stats"><div class="stat"><strong>~46%</strong><span>exact correction</span></div><div class="stat"><strong>69%</strong><span>corrupted examples improved</span></div><div class="stat"><strong>-49%</strong><span>total edit distance</span></div><div class="stat"><strong>~6%</strong><span>examples worsened</span></div></div>
-    <div class="callout"><strong>Why this matters</strong><p>The result suggests that pretrained Eris representations can transfer to a new correction task with very little supervised data, even without synthetic corruption examples.</p></div>
-  </div></section>
+    <div class="stats">
+      <div class="stat"><strong>500</strong><span>manually corrected real-world pairs</span></div>
+      <div class="stat"><strong>~46%</strong><span>exact correction</span></div>
+      <div class="stat"><strong>69%</strong><span>corrupted examples improved</span></div>
+      <div class="stat"><strong>-49%</strong><span>total edit distance</span></div>
+    </div>
 
-  <section><div class="wrap">
-    <div class="eyebrow">Current limitations</div><h3 class="section-title">Calibration and cheaper global context are now the main targets.</h3>
-    <div class="prose"><p>Current limitations are mainly <strong>edit-decision calibration</strong>, insert/delete handling, protection of already-correct text, and the compute cost of the compressed global pathway.</p><p>The next architecture iteration is focused on making C4 global-context injection substantially cheaper while preserving the reconstruction benefit already confirmed by ablation.</p></div>
+    <div class="table-wrap"><table>
+      <thead><tr><th>Fine-tuning result</th><th>Held-out result</th></tr></thead>
+      <tbody>
+        <tr class="highlight"><td>Exact correction</td><td>≈ 46%</td></tr>
+        <tr><td>Corrupted examples improved</td><td>69%</td></tr>
+        <tr><td>Total edit-distance change</td><td>−49%</td></tr>
+        <tr><td>Examples worsened</td><td>≈ 6%</td></tr>
+      </tbody>
+    </table></div>
+
+    <div class="callout">
+      <strong>Current limitations</strong>
+      <p>
+        The main weaknesses are <strong>edit-decision calibration</strong>,
+        <strong>insert/delete handling</strong>, protection of already-correct text, and the compute
+        cost of the global pathway. Ongoing work is focused on these issues before scaling Eris further.
+      </p>
+    </div>
+
+    <p class="note">
+      These results remain preliminary. They demonstrate stable billion-token-scale training,
+      useful global-context learning and encouraging low-data downstream transfer, but they are not
+      a finished-model quality comparison against established pretrained language models.
+    </p>
   </div></section>
 
   <section><div class="wrap">
@@ -289,6 +365,11 @@ function showIndex(){
   history.replaceState(null,"","#models");
   window.scrollTo(0,0);
 }
+function scrollInsideModel(targetId){
+  const target=document.getElementById(targetId);
+  if(!target) return;
+  target.scrollIntoView({behavior:"smooth",block:"start"});
+}
 window.addEventListener("DOMContentLoaded",()=>{
   const h=location.hash.slice(1);
   if(ids.includes(h)) showModel(h); else showIndex();