Open-source, commercial-use-friendly brain-MR multimodal reasoning model.
medimage-model is the fully-permissive track of the GOATnote medical imaging stack. Trained only on commercial-use-friendly data (OpenNeuro CC0, UCSF-BMSR CC BY 4.0, open chest corpora, NVIDIA Open Model License synthesis) with the medimage-eval substrate as the evaluation contract.
Status: pre-v0.1 scaffold. The research-track sibling with MR-RATE training is
medimage-model-research— published in parallel under the same substrate, different data-license boundary.
Most large open medical-imaging paired-report corpora (MR-RATE, CT-RATE) are CC BY-NC-SA. Models trained on them carry an unclear-but-cautious "research-only" cloud. This repo deliberately avoids that contamination: every dataset entering training is verified permissive at ingest, and provenance is hashed into the model card.
- Image encoder: MAISI-v2 latents (NVIDIA Open Model License, commercial-OK) for brain MR; MedImageInsight / MedSAM2 backbones for other modalities.
- Text decoder: Nemotron-3-Nano-30B-A3B-BF16 (NVIDIA Open Model License) via PEFT/LoRA on Mamba-2 + MoE projections.
- Synthesis engine: NV-Generate-MR-Brain (NVIDIA Open Model License, royalty-free fine-tune) for long-tail augmentation; strata caps + per-sample tagging.
- Training: Megatron + SGLang + GRPO (reused from
healthcraftRL coupling). - Eval:
medimage-evalsubstrate (dual-judge κ + shift gauntlet + adjudication + receipts).
pip install medimage-modelFrom source:
pip install git+https://github.com/GOATnote-Inc/medimage-model.git@mainCode: Apache License 2.0. Trained weights: NVIDIA Open Model License (commercial-OK) — released per release in releases/vX.Y/model_card.md. Training data licenses: see docs/LICENSE_MAP.md. No CC BY-NC-SA contamination.