Apppublic

lablab-ai-amd-developer-hackathon/lysos

sourceHugging Facemitupdated 4mo agoView on Hugging Face

2likes

App README

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

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

📂 Full source: github.com/Rahul-Rajpurohitk/lysos