Static landing page · Lysos · MI300X + Gemma 4

README.md

@@ -1,10 +1,104 @@
 ---
 title: Lysos
-emoji: 📈
-colorFrom: pink
-colorTo: blue
-sdk: docker
-pinned: false
+emoji: 🧬
+colorFrom: purple
+colorTo: red
+sdk: static
+pinned: true
+license: mit
+short_description: Agentic antibiotic designer · Gemma 4 + MI300X
+tags:
+  - antibiotic-discovery
+  - drug-design
+  - gemma
+  - mi300x
+  - amd
+  - amr
+  - agentic-ai
+  - antimicrobial-resistance
 ---
 
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+# Lysos · Open-source antibiotic designer for the AMR pandemic
+
+> **Three-stage fine-tune of Gemma 4 31B-it on AMD MI300X. Multi-agent debate engine. End-to-end live agentic workspace.**
+
+🏆 **AMD Developer Hackathon 2026** · Track 2 — Fine-Tuning on AMD GPUs
+
+---
+
+## What it is
+
+Lysos is an end-to-end open-source antibiotic discovery platform that takes Google's **Gemma 4 31B-it** and specializes it for antimicrobial-resistance (AMR) drug design via a three-stage fine-tune on a single AMD MI300X. The fine-tuned model drives a multi-agent debate engine and a live agentic workspace.
+
+## Live links
+
+| Asset | Where |
+|---|---|
+| 📂 GitHub repo (full source) | <https://github.com/Rahul-Rajpurohitk/lysos> |
+| 🤖 Stage 2.5 model (production) | <https://huggingface.co/rahul24raj/lysos-base-dpo> |
+| 🤖 Stage 2 model | <https://huggingface.co/rahul24raj/lysos-base> |
+| 🤖 Stage 1 model | <https://huggingface.co/rahul24raj/txgemma-4-31b> |
+| 📊 Stage 2 SFT dataset (222,606 AMR examples) | <https://huggingface.co/datasets/rahul24raj/lysos-amr-stage2> |
+| 🎬 Demo videos (release) | <https://github.com/Rahul-Rajpurohitk/lysos/releases/tag/v1.0-hackathon-submission> |
+| 📺 Full walkthrough (9:08) | [lysos-demo-merged.mp4](https://github.com/Rahul-Rajpurohitk/lysos/releases/download/v1.0-hackathon-submission/lysos-demo-merged.mp4) |
+
+## The three-stage fine-tune
+
+Every stage trains a LoRA adapter on top of `google/gemma-4-31B-it`. All adapters are public.
+
+```
+                          google/gemma-4-31B-it (62 GB base)
+                                       │
+       ┌───────────────────────────────┼───────────────────────────────┐
+       │                               │                               │
+   STAGE 1                          STAGE 2                        STAGE 2.5
+TxGemma-4 31B                  lysos-base                   lysos-base-dpo
+LoRA r=64, α=256              LoRA r=64, α=128             LoRA r=32, α=64 (β=0.1)
+continued pretraining         SFT on 222,606 AMR examples  DPO on hard-negative pairs
+for therapeutics              (8 priority pathogens)       (10 anti-correlated axes)
+~2 hr on 1× MI300X            ~3 hr on 1× MI300X           ~45 min on 1× MI300X
+```
+
+**Why DPO for the alignment stage**: DPO is the right tool for this objective. The downstream usage pattern — the Strategist agent picking among Designer-proposed candidates — is a discrete preference choice, exactly what DPO optimizes for. KL-bounded objective for stability, no axis to game, sample-efficient at 10K pairs in 45 min on 1× MI300X, full base capability preserved.
+
+## The agentic workspace
+
+When you fire `/wf design_with_debate`, four agent roles take turns — each is a separate LLM call:
+
+```
+  ┌───────────┐         ┌───────────┐         ┌───────────┐         ┌─────────────┐
+  │ DESIGNER  │── 3 ───▶│  CRITIC   │──────▶─▶│  EDITOR   │──────▶─▶│  STRATEGIST │
+  │  drafts   │ smiles  │ challenges│ critique│ refines   │  fix    │  picks winner│
+  └───────────┘         └───────────┘         └───────────┘         └──────┬──────┘
+                                                                           │
+                                            winner SMILES auto-loads to 2D + 3D + radar
+```
+
+Plus 7 streaming workflows, 12+ slash commands, real-time per-atom resistance scoring against curated CARD clinical mutations, per-pathogen Champion table, Knowledge command-center with 4-tier resistance gene network.
+
+## Why MI300X
+
+192 GB HBM3 lets us fit **Gemma 4 31B base in bf16 + LoRA adapter + KV cache + agent context coresident on one GPU**. Same GPU trains and serves. No tensor parallelism, no model sharding, no migration step.
+
+## Run it locally
+
+```bash
+git clone https://github.com/Rahul-Rajpurohitk/lysos.git
+cd lysos
+python3 -m venv .venv && source .venv/bin/activate
+pip install -e .
+uvicorn workspace.api.server:app --host 0.0.0.0 --port 7860 &
+
+cd workspace/web && npm install && npm run dev
+# open http://localhost:5173
+```
+
+## License
+
+MIT (code) · Apache-2.0 / Gemma terms (weights) · CC-BY (datasets)
+
+---
+
+📺 **Watch the 9-minute demo**: [lysos-demo-merged.mp4](https://github.com/Rahul-Rajpurohitk/lysos/releases/download/v1.0-hackathon-submission/lysos-demo-merged.mp4)
+
+📂 **Full source**: [github.com/Rahul-Rajpurohitk/lysos](https://github.com/Rahul-Rajpurohitk/lysos)

index.html

@@ -0,0 +1,248 @@
+<!DOCTYPE html>
+<html lang="en">
+<head>
+  <meta charset="utf-8" />
+  <meta name="viewport" content="width=device-width, initial-scale=1" />
+  <title>Lysos · Open-source antibiotic designer · Gemma 4 + AMD MI300X</title>
+  <meta name="description" content="Three-stage fine-tune of Gemma 4 31B-it on AMD MI300X. Multi-agent debate engine. End-to-end live agentic workspace for antimicrobial-resistance drug design." />
+  <style>
+    :root {
+      --bg-0: #0a0a14;
+      --bg-1: #1a1a2e;
+      --purple: #8458ff;
+      --purple-2: #c2adff;
+      --red: #ED1C24;
+      --green: #39e08e;
+      --amber: #f59e0b;
+      --text: #e6ecff;
+      --text-dim: #9aa3b8;
+      --text-faint: #64748b;
+      --border: rgba(132,88,255,0.18);
+    }
+    * { box-sizing: border-box; }
+    body {
+      margin: 0;
+      font-family: -apple-system, BlinkMacSystemFont, "Inter", "SF Pro", system-ui, sans-serif;
+      background: linear-gradient(180deg, var(--bg-0), var(--bg-1));
+      color: var(--text);
+      min-height: 100vh;
+      line-height: 1.6;
+    }
+    .wrap { max-width: 880px; margin: 0 auto; padding: 60px 24px 80px; }
+    .hero { text-align: center; margin-bottom: 56px; }
+    h1 {
+      font-size: 56px; margin: 0 0 12px;
+      background: linear-gradient(135deg, var(--purple-2), var(--purple), var(--red));
+      -webkit-background-clip: text; background-clip: text; color: transparent;
+      letter-spacing: -0.02em; font-weight: 800;
+    }
+    .tagline { font-size: 18px; color: var(--text-dim); max-width: 640px; margin: 0 auto 20px; }
+    .badges { display: flex; justify-content: center; gap: 8px; flex-wrap: wrap; margin-top: 18px; }
+    .badge {
+      padding: 5px 12px; font-size: 11px; font-weight: 700;
+      border-radius: 999px; letter-spacing: 0.04em; text-transform: uppercase;
+      background: rgba(132,88,255,0.12); color: var(--purple-2); border: 1px solid var(--border);
+    }
+    .badge.amd { background: rgba(237,28,36,0.10); color: #ff6b6b; border-color: rgba(237,28,36,0.30); }
+    .badge.green { background: rgba(57,224,142,0.10); color: var(--green); border-color: rgba(57,224,142,0.30); }
+    .grid {
+      display: grid; gap: 14px; grid-template-columns: repeat(2, 1fr);
+      margin-bottom: 40px;
+    }
+    @media (max-width: 680px) { .grid { grid-template-columns: 1fr; } }
+    .card {
+      padding: 18px 20px;
+      background: rgba(132,88,255,0.04);
+      border: 1px solid var(--border); border-left: 3px solid var(--purple);
+      border-radius: 8px;
+    }
+    .card a { color: var(--purple-2); text-decoration: none; font-weight: 600; }
+    .card a:hover { color: var(--purple); }
+    .card .label {
+      font-size: 9.5px; color: var(--text-faint); text-transform: uppercase;
+      letter-spacing: 1px; font-weight: 700; margin-bottom: 4px;
+    }
+    .card h3 { margin: 0 0 4px; font-size: 16px; color: var(--text); font-weight: 700; }
+    .card p { margin: 0; font-size: 13px; color: var(--text-dim); }
+    h2 {
+      font-size: 26px; margin: 56px 0 16px; color: var(--text); font-weight: 700;
+      letter-spacing: -0.01em;
+    }
+    h2:before { content: "▸ "; color: var(--purple); margin-right: 4px; }
+    pre {
+      background: rgba(0,0,0,0.30); border: 1px solid var(--border);
+      padding: 16px 20px; border-radius: 6px;
+      font-family: "JetBrains Mono", "Fira Code", monospace;
+      font-size: 12px; color: var(--purple-2); line-height: 1.6;
+      overflow-x: auto;
+    }
+    .stages {
+      display: grid; grid-template-columns: repeat(3, 1fr); gap: 10px;
+      margin: 24px 0;
+    }
+    @media (max-width: 680px) { .stages { grid-template-columns: 1fr; } }
+    .stage {
+      padding: 16px; background: rgba(132,88,255,0.06);
+      border: 1px solid var(--border); border-radius: 6px;
+    }
+    .stage .num {
+      font-size: 10px; color: var(--purple); font-weight: 800; letter-spacing: 1.5px;
+    }
+    .stage .name {
+      font-family: "JetBrains Mono", monospace; font-size: 13px;
+      color: var(--text); font-weight: 700; margin: 4px 0;
+    }
+    .stage .meta { font-size: 11px; color: var(--text-faint); }
+    .video-card {
+      background: linear-gradient(135deg, rgba(237,28,36,0.06), rgba(132,88,255,0.06));
+      border: 1px solid rgba(237,28,36,0.30);
+      padding: 24px; border-radius: 10px; text-align: center;
+      margin: 32px 0;
+    }
+    .video-card a {
+      display: inline-block; margin-top: 8px; padding: 10px 22px;
+      background: var(--red); color: white;
+      border-radius: 6px; text-decoration: none; font-weight: 700;
+      letter-spacing: 0.04em;
+    }
+    footer { margin-top: 60px; text-align: center; color: var(--text-faint); font-size: 12px; }
+    footer a { color: var(--purple-2); text-decoration: none; }
+    .agent-flow {
+      display: flex; align-items: center; justify-content: center;
+      gap: 6px; flex-wrap: wrap; margin: 18px 0;
+    }
+    .agent {
+      padding: 8px 14px; border-radius: 999px;
+      font-family: "JetBrains Mono", monospace; font-size: 11px; font-weight: 700;
+      letter-spacing: 0.04em;
+    }
+    .a-d { background: rgba(57,224,142,0.10); color: var(--green); border: 1px solid rgba(57,224,142,0.40); }
+    .a-c { background: rgba(237,28,36,0.10); color: #ff6b6b; border: 1px solid rgba(237,28,36,0.40); }
+    .a-e { background: rgba(93,138,255,0.10); color: #5d8aff; border: 1px solid rgba(93,138,255,0.40); }
+    .a-s { background: rgba(132,88,255,0.10); color: var(--purple-2); border: 1px solid rgba(132,88,255,0.40); }
+    .arr { color: var(--text-faint); font-family: "JetBrains Mono", monospace; }
+  </style>
+</head>
+<body>
+<div class="wrap">
+
+  <div class="hero">
+    <h1>Lysos</h1>
+    <div class="tagline">
+      Open-source generative antibiotic designer for the AMR pandemic.<br/>
+      Three-stage fine-tune of Gemma 4 31B-it on AMD MI300X. Multi-agent debate engine. End-to-end live agentic workspace.
+    </div>
+    <div class="badges">
+      <span class="badge amd">🏆 AMD Developer Hackathon 2026</span>
+      <span class="badge">Track 2 · Fine-Tuning on AMD GPUs</span>
+      <span class="badge green">MIT License</span>
+    </div>
+  </div>
+
+  <h2>Live links</h2>
+  <div class="grid">
+    <div class="card">
+      <div class="label">Source</div>
+      <h3><a href="https://github.com/Rahul-Rajpurohitk/lysos">GitHub repo →</a></h3>
+      <p>FastAPI backend · React+Vite frontend · training pipeline · agentic harness · MIT</p>
+    </div>
+    <div class="card">
+      <div class="label">Production model</div>
+      <h3><a href="https://huggingface.co/rahul24raj/lysos-base-dpo">rahul24raj/lysos-base-dpo →</a></h3>
+      <p>Stage 2.5 DPO · LoRA r=32 · adapter on Gemma 4 31B-it</p>
+    </div>
+    <div class="card">
+      <div class="label">SFT model</div>
+      <h3><a href="https://huggingface.co/rahul24raj/lysos-base">rahul24raj/lysos-base →</a></h3>
+      <p>Stage 2 SFT · LoRA r=64 · 222,606 AMR examples</p>
+    </div>
+    <div class="card">
+      <div class="label">Pretraining model</div>
+      <h3><a href="https://huggingface.co/rahul24raj/txgemma-4-31b">rahul24raj/txgemma-4-31b →</a></h3>
+      <p>Stage 1 · continued pretraining for therapeutics · LoRA r=64 α=256</p>
+    </div>
+    <div class="card">
+      <div class="label">Training dataset</div>
+      <h3><a href="https://huggingface.co/datasets/rahul24raj/lysos-amr-stage2">lysos-amr-stage2 →</a></h3>
+      <p>222,606 AMR records · 8 priority pathogens · ChEMBL + literature + curated negatives</p>
+    </div>
+    <div class="card">
+      <div class="label">Demo videos</div>
+      <h3><a href="https://github.com/Rahul-Rajpurohitk/lysos/releases/tag/v1.0-hackathon-submission">v1.0 release →</a></h3>
+      <p>Merged 9:08 walkthrough + 3 individual scenes (agentic flow, system tour, full)</p>
+    </div>
+  </div>
+
+  <div class="video-card">
+    <div class="label" style="color: var(--text-dim); font-size: 10px; letter-spacing: 1px; text-transform: uppercase; font-weight: 700;">9-MINUTE FULL WALKTHROUGH</div>
+    <h3 style="margin: 8px 0; font-size: 22px; color: var(--text);">Watch the demo</h3>
+    <p style="margin: 0 0 12px; color: var(--text-dim); font-size: 13px;">
+      Multi-agent debate · live workspace · real-time scoring · resistance escape map · Pareto frontier
+    </p>
+    <a href="https://github.com/Rahul-Rajpurohitk/lysos/releases/download/v1.0-hackathon-submission/lysos-demo-merged.mp4">▶ Play (41 MB)</a>
+  </div>
+
+  <h2>The three-stage fine-tune</h2>
+  <p style="color: var(--text-dim);">Every stage trains a LoRA adapter on top of <code style="color: var(--purple-2);">google/gemma-4-31B-it</code>. All adapters are public on Hugging Face.</p>
+  <div class="stages">
+    <div class="stage">
+      <div class="num">STAGE 1</div>
+      <div class="name">TxGemma-4 31B</div>
+      <div class="meta">LoRA r=64, α=256 · ~2 hr on 1× MI300X<br/>Continued pretraining for therapeutics</div>
+    </div>
+    <div class="stage">
+      <div class="num">STAGE 2</div>
+      <div class="name">lysos-base</div>
+      <div class="meta">LoRA r=64, α=128 · ~3 hr on 1× MI300X<br/>SFT on 222,606 AMR examples</div>
+    </div>
+    <div class="stage">
+      <div class="num">STAGE 2.5</div>
+      <div class="name">lysos-base-dpo</div>
+      <div class="meta">LoRA r=32, α=64, β=0.1 · ~45 min on 1× MI300X<br/>DPO on 10K hard-negative Pareto pairs</div>
+    </div>
+  </div>
+
+  <h2>The multi-agent debate</h2>
+  <p style="color: var(--text-dim);">When you fire <code style="color: var(--purple-2);">/wf design_with_debate</code>, four agent roles take turns — each is a separate LLM call.</p>
+  <div class="agent-flow">
+    <span class="agent a-d">DESIGNER</span>
+    <span class="arr">─ proposes 3 ──▶</span>
+    <span class="agent a-c">CRITIC</span>
+    <span class="arr">── critiques ──▶</span>
+    <span class="agent a-e">EDITOR</span>
+    <span class="arr">── refines ──▶</span>
+    <span class="agent a-s">STRATEGIST</span>
+  </div>
+  <p style="color: var(--text-dim); font-size: 13px; text-align: center; margin-top: 14px;">
+    Strategist picks winner + runner-up + next action.<br/>
+    Winner SMILES auto-loads to 2D builder + 3D pocket viewer + 12-axis reward radar + resistance escape map.
+  </p>
+
+  <h2>Why MI300X is load-bearing</h2>
+  <p style="color: var(--text-dim);">
+    192 GB HBM3 lets us fit Gemma 4 31B base in bf16 + LoRA adapter + KV cache + agent context
+    <strong style="color: var(--text);">coresident on one GPU</strong>. Same GPU trains and serves.
+    No tensor parallelism, no model sharding, no migration step.
+    Total wall-clock for the three-stage fine-tune: <strong style="color: var(--purple-2);">~6 hours on 1 GPU</strong>.
+  </p>
+
+  <h2>Run it locally</h2>
+  <pre>git clone https://github.com/Rahul-Rajpurohitk/lysos.git
+cd lysos
+python3 -m venv .venv && source .venv/bin/activate
+pip install -e .
+uvicorn workspace.api.server:app --host 0.0.0.0 --port 7860 &amp;
+
+cd workspace/web && npm install && npm run dev
+# open http://localhost:5173</pre>
+
+  <footer>
+    <p>
+      <strong>License</strong>: MIT (code) · Apache-2.0 / Gemma terms (weights) · CC-BY (datasets)<br/>
+      Built for the <a href="https://lablab.ai">AMD Developer Hackathon 2026</a> · Track 2: Fine-Tuning on AMD GPUs
+    </p>
+  </footer>
+
+</div>
+</body>
+</html>