Pixel-art scene — Stanley, standing at a tall wooden card-catalog with a wooden cane in one hand, files a softly glowing record card into a drawer; a warm lamp, a potted plant, and a deep cosmic-purple window behind
Rare Disease Active

Rare AI Archive

A curated, open-weight language model and evaluation toolkit fine-tuned on rare disease literature — from diagnostic reasoning to patient advocacy.

rare-diseaselanguage-modelsopen-weightfine-tuning

The Rare AI Archive hosts open-weight models trained on both rare disease research and patient experience. Every model ships with a lineage card tracing its training data and evaluation metrics.

Built with the Wilhelm Foundation, the archive is a complete open-source rare disease diagnostic AI toolkit on HuggingFace. The centerpiece is rare-archive-qwen-4b-sft-v1 — a 4-billion parameter language model fine-tuned on rare disease clinical reasoning. Three companion evaluation datasets (RareArena-RDS, RareArena-RDC, and synthetic patients) provide standardized benchmarks, and a clinical demo Space lets physicians test diagnostic reasoning against ten real-world scenarios.

Under the hood, that centerpiece model is a QLoRA fine-tune of Qwen3.5-4B, trained on 63,212 rare-disease reasoning examples. On a held-out test set it lifts top-1 diagnostic accuracy from about 1% to 21.5% — a real jump, and deliberately an honest one: this is Stage 1 of a planned multi-stage pipeline, released open (Apache-2.0) precisely so its limits are inspectable. It’s a research and decision-support tool, not a medical device, and not a substitute for a clinician. A closed-source system or two still scores higher on raw benchmarks; the Archive trades a few points of peak accuracy for something the closed models can’t offer — weights anyone can download, audit, and build on.

The tagline says it: “No disease is too rare to matter.” Open weights mean any researcher, anywhere, can download the model and run it on their own infrastructure. No API keys. No usage limits. No gatekeepers between a clinician’s question and a tool that might help answer it.

Behind every dataset is a patient. Behind every model is hope. The Rare AI Archive is what happens when you take that seriously and ship the infrastructure to match.