Medical AI, in the open.
The models and datasets I publish on Hugging Face. It starts with an embedding model that searches Japanese medical text. That search then builds the before/after guideline datasets, which lead up to a completion model for writing clinical notes. The manga speech-bubble OCR is where I took the small-model training I'd picked up along the way and tried it outside medicine.
Baberu OCR
Apache-2.0Reads manga speech bubbles with 115M parameters. Handles Japanese, Chinese and English; on Japanese it beats the teacher model it was distilled from, and stays competitive with models 8× its size at the bubble level.
med-ruri v3
EmbeddingsAn embedding model for searching Japanese medical text. ModernBERT-Ja, further trained on the medical domain. Seven variants from 70M to 310M, used as the retrieval backbone for in-house RAG and for building the datasets above.
med-lfm2.5 autocomplete
1.2BA completion model that predicts what comes next as you write charts and discharge summaries. Built on LiquidAI's LFM2.5, with four stages of medical fine-tuning. Light enough to run offline on CPU.
med-slm-ja Before / After
CC BY 4.0Traces revisions across 201 Japanese clinical guidelines (2010–2024), recording how the answer to the same question changed between the old and new editions, with citations for both. 46,705 pairs, including 7,042 where the answer changed clinically. There's a live demo.
med-guideline-temporal-en
CC BY-NC 4.0The English counterpart of the dataset above. It follows revisions across 19 US medical societies and 121 series, structuring how recommendations changed. For training time-aware retrieval, and for measuring how fast knowledge goes stale.