
Lysos is an end-to-end agentic antibiotic discovery platform. We took Google's Gemma 4 31B-it and fine-tuned it three times on AMD MI300X: Stage 1 — Continued pretraining for therapeutics (TxGemma-4 31B) Stage 2 — Supervised fine-tuning on 222,606 AMR examples (lysos-base) Stage 2.5 — DPO alignment on hard-negative Pareto pairs (lysos-base-dpo) The chat-driven workspace runs a multi-agent debate (Designer / Critic / Editor / Strategist), an orchestrator agent that routes free text through 7 streaming workflows + 12 slash commands, a per-pathogen Champion table, a Knowledge command-center with resistance-gene → drug-class network mapping, and a Pareto frontier explorer with Critic narration. Every agent action is recorded with token cost, p50/p95/p99 latency, and cross-agent handoff edges. All training artifacts are public on Hugging Face. The merged model serves via vllm on AMD MI300X behind an OpenAI-compatible endpoint. Repo + HF visibility status: - ✓ GitHub Rahul-Rajpurohitk/lysos PUBLIC - ⏳ HF models — you need to flip the 4 visibility toggles linked above
10 May 2026