Upload index.html

index.html

@@ -63,6 +63,14 @@ tr.highlight{background:var(--soft)}
 .version-card{border:1px solid var(--line);background:var(--panel);padding:20px}
 .version-card small{color:var(--accent);font-size:10px;text-transform:uppercase;letter-spacing:.1em}
 .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-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}
+@media(max-width:900px){.update-item{grid-template-columns:1fr;gap:5px}}
+
 footer{border-top:1px solid var(--line);padding:36px 0;color:var(--muted);font-size:12px}
 @media(max-width:900px){.model-grid,.stats,.version-grid{grid-template-columns:1fr}.timeline-item{grid-template-columns:1fr;gap:8px}.nav-note{display:none}}
 </style>
@@ -112,7 +120,7 @@ footer{border-top:1px solid var(--line);padding:36px 0;color:var(--muted);font-s
 
 <article class="detail" id="model-polmath">
   <div class="detail-top"><div class="wrap detail-bar"><button class="back" onclick="showIndex()">← All models</button><div class="detail-name">PolMATH</div></div></div>
-  <div class="wrap detail-hero"><div class="eyebrow">Numerical representation research</div><h2>PolMATH</h2><p>A custom model exploring numerical value as a structured signal separate from the surface text used to write it.</p></div>
+  <div class="wrap detail-hero"><div class="eyebrow">Numerical representation research</div><h2>PolMATH</h2><p>A custom model exploring numerical value as a structured signal separate from the surface text used to write it.</p><div class="update-log"><div class="update-item"><div class="update-date">2025</div><div class="update-copy"><strong>Initial architecture and training experiments</strong><span>Structured numerical channel, numeric decoding and number↔text association tests.</span></div></div><div class="update-item"><div class="update-date">Current</div><div class="update-copy"><strong>Publication-oriented rework</strong><span>Original branch paused; cleaner protocol and English-language continuation are being prepared.</span></div></div></div></div>
   <section><div class="wrap">
     <div class="eyebrow">Core idea</div><h3 class="section-title">Numbers were not treated as ordinary text alone.</h3>
     <div class="prose">
@@ -136,7 +144,7 @@ footer{border-top:1px solid var(--line);padding:36px 0;color:var(--muted);font-s
   <div class="wrap detail-hero">
     <div class="eyebrow">Structural compression and recovery</div><h2>ORIS 660M</h2>
     <p>A 660,131,328-parameter student derived from Bielik-1.5B-v3 by collapsing 32 transformer blocks into a selected 12-block path while preserving the teacher's width and core geometry.</p>
-    <div class="stats"><div class="stat"><strong>660.13M</strong><span>parameters</span></div><div class="stat"><strong>32 → 12</strong><span>blocks</span></div><div class="stat"><strong>621.984M</strong><span>recovery labels</span></div><div class="stat"><strong>Paused</strong><span>research active</span></div></div>
+    <div class="stats"><div class="stat"><strong>660.13M</strong><span>parameters</span></div><div class="stat"><strong>32 → 12</strong><span>blocks</span></div><div class="stat"><strong>621.984M</strong><span>recovery labels</span></div><div class="stat"><strong>Paused</strong><span>research active</span></div></div><div class="update-log"><div class="update-item"><div class="update-date">07 Aug 2026</div><div class="update-copy"><strong>ORIS 660M branch begins</strong><span>Initial structural-reduction experiment from Bielik-1.5B-v3.</span></div></div><div class="update-item"><div class="update-date">Recovery tests</div><div class="update-copy"><strong>Selected-vs-uniform architecture comparison</strong><span>Confirmed that retained layer identity materially changes recovery trajectory.</span></div></div><div class="update-item"><div class="update-date">Later branch</div><div class="update-copy"><strong>Knowledge and data-control experiments</strong><span>Inherited-vs-random controls, diagnostic benchmarks and planned staged recovery curriculum.</span></div></div></div>
   </div>
   <section><div class="wrap">
     <div class="eyebrow">Versions / experiments</div><h3 class="section-title">The project is a sequence of recovery experiments.</h3>
@@ -167,7 +175,7 @@ footer{border-top:1px solid var(--line);padding:36px 0;color:var(--muted);font-s
   <div class="wrap detail-hero">
     <div class="eyebrow">Compact Polish encoder</div><h2>ORIS Small C</h2>
     <p>Originally developed under the working name ORIS Bert Small C. It is a custom encoder using BERT-style MLM training, not a stock reduced BERT.</p>
-    <div class="stats"><div class="stat"><strong>25.41M</strong><span>parameters</span></div><div class="stat"><strong>6</strong><span>layers</span></div><div class="stat"><strong>128K</strong><span>BPE vocab</span></div><div class="stat"><strong>8.00B</strong><span>pretraining tokens</span></div></div>
+    <div class="stats"><div class="stat"><strong>25.41M</strong><span>parameters</span></div><div class="stat"><strong>6</strong><span>layers</span></div><div class="stat"><strong>128K</strong><span>BPE vocab</span></div><div class="stat"><strong>8.00B</strong><span>pretraining tokens</span></div></div><div class="update-log"><div class="update-item"><div class="update-date">16 Aug 2026</div><div class="update-copy"><strong>Small C checkpoint completed</strong><span>8B-token from-scratch MLM pretraining completed.</span></div></div><div class="update-item"><div class="update-date">Downstream tests</div><div class="update-copy"><strong>Filtering and encoder-efficiency evaluation</strong><span>Pipeline, KLEJ-style and local throughput tests established the compact encoder's practical role.</span></div></div></div>
   </div>
   <section><div class="wrap">
     <div class="eyebrow">Architecture</div><h3 class="section-title">Factorized embeddings, RMSNorm and mixed local/global attention.</h3>
@@ -185,7 +193,7 @@ footer{border-top:1px solid var(--line);padding:36px 0;color:var(--muted);font-s
 
 <article class="detail" id="model-cain">
   <div class="detail-top"><div class="wrap detail-bar"><button class="back" onclick="showIndex()">← All models</button><div class="detail-name">CAIN</div></div></div>
-  <div class="wrap detail-hero"><div class="eyebrow">Persona research</div><h2>CAIN</h2><p>An unpublished Bielik 1.5B fine-tune exploring strong persona conditioning and behavioural dependence on SFT-defined identity.</p></div>
+  <div class="wrap detail-hero"><div class="eyebrow">Persona research</div><h2>CAIN</h2><p>An unpublished Bielik 1.5B fine-tune exploring strong persona conditioning and behavioural dependence on SFT-defined identity.</p><div class="update-log"><div class="update-item"><div class="update-date">2026</div><div class="update-copy"><strong>Persona-focused SFT experiment</strong><span>Strong conversational identity, architecture-awareness data and light post-SFT reinforcement work.</span></div></div><div class="update-item"><div class="update-date">Research hold</div><div class="update-copy"><strong>Broader persona-study concept</strong><span>Not published while the experiment is reframed around persona persistence and behavioural dependence.</span></div></div></div></div>
   <section><div class="wrap">
     <div class="eyebrow">Research direction</div><h3 class="section-title">Persona as something deeper than writing style.</h3>
     <div class="prose"><p>CAIN was trained with a deliberately opinionated, informal persona, light RL and SFT data that included awareness of its own architecture. It tended toward antihero choices, self-sacrifice, unusual trade-offs and dry humour.</p><p>The project remains unpublished because the more interesting direction is a broader study of persona persistence and behavioural dependence in LLMs.</p></div>
@@ -197,14 +205,14 @@ 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="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>
 
   <section><div class="wrap">
     <div class="eyebrow">Architecture</div><h3 class="section-title">Parallel denoising instead of sequential next-token generation.</h3>
     <div class="prose">
       <p>Eris uses a full-resolution token pathway together with a compressed global C4 latent pathway. Training uses denoising reconstruction plus an auxiliary corruption-detection objective.</p>
-      <p>The current prototype was initialized from scratch and has progressed beyond inference-only smoke tests into billion-token-scale Polish foundation pretraining.</p>
+      <p>The current prototype was initialized from scratch and has progressed beyond inference-only smoke tests into billion-token-scale Polish foundation pretraining. At the reported ~2.8B-token stage it trained at roughly <strong>80k tokens/s on an RTX 5060 Ti</strong> and used a ~32k Bielik/APT4 tokenizer.</p>
     </div>
   </div></section>
 
@@ -239,13 +247,26 @@ footer{border-top:1px solid var(--line);padding:36px 0;color:var(--muted);font-s
     <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>
 
+  <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>
+
+  <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></section>
+
   <section><div class="wrap">
     <div class="eyebrow">Roadmap</div><h3 class="section-title">From architecture smoke test to foundation model.</h3>
     <div class="timeline">
-      <div class="timeline-item"><div class="timeline-year">Phase 1</div><div class="timeline-body"><h4>Inference prototype</h4><p>Synthetic edit workloads established the batching and parallel-refinement throughput signal.</p></div></div>
-      <div class="timeline-item"><div class="timeline-year">Phase 2</div><div class="timeline-body"><h4>~43M foundation prototype</h4><p>Random initialization and billion-token Polish pretraining with denoising reconstruction.</p></div></div>
-      <div class="timeline-item"><div class="timeline-year">Now</div><div class="timeline-body"><h4>Global-context efficiency</h4><p>Reduce the compute cost of C4 without losing the reconstruction benefit demonstrated by ablation.</p></div></div>
-      <div class="timeline-item"><div class="timeline-year">Next</div><div class="timeline-body"><h4>Objective and geometry</h4><p>Refine reconstruction behaviour, training objective, final model geometry and direct quality evaluation.</p></div></div>
+      <div class="timeline-item"><div class="timeline-year">18 Aug 2026</div><div class="timeline-body"><h4>Architecture benchmark stage</h4><p>Synthetic edit workloads established the batching and parallel-refinement throughput signal.</p></div></div>
+      <div class="timeline-item"><div class="timeline-year">19 Aug 2026</div><div class="timeline-body"><h4>First pretrained foundation prototype</h4><p>~43M parameters, ~32k Bielik/APT4 tokenizer, random initialization and ~2.8B Polish-language training tokens at the reported evaluation.</p></div></div>
+      <div class="timeline-item"><div class="timeline-year">19 Aug 2026</div><div class="timeline-body"><h4>First downstream fine-tuning result</h4><p>500 manually corrected real-world pairs produced ~46% exact correction, 69% improvement rate and 49% lower total edit distance.</p></div></div>
+      <div class="timeline-item"><div class="timeline-year">Current</div><div class="timeline-body"><h4>Calibration and global-context efficiency</h4><p>Improve edit decisions, insert/delete handling and protection of correct text while reducing C4 compute cost.</p></div></div>
+      <div class="timeline-item"><div class="timeline-year">Next</div><div class="timeline-body"><h4>Scale only after architecture cleanup</h4><p>Refine reconstruction behaviour, objective and final model geometry before larger-scale training and direct quality comparisons.</p></div></div>
     </div>
   </div></section>
 </article>