med-slm-ja Before / After

What set this off was watching models answer with old knowledge, full of confidence. A model doesn't tell which point in time its recommendation is from. So — make the moment the answer changed the data itself. Put old and new side by side and you get both a yardstick for how stale knowledge is, and teaching material for training time-aware retrieval.

In medicine this is life-and-death. When to start an anticoagulant, first-line cancer therapy — the standard flipping within a few years is nothing unusual. That's why it was the first subject I picked.

First I raise a Q&A from the new edition, then retrieve the matching old-edition text with med-ruri's embedding search. Gemma writes both editions' answers from the passages retrieved and sorts the difference into five types. Things embedding search tends to miss — a new drug, say — are picked up separately by keyword. The retrieval backbone is my own med-ruri, used as-is. Feeding one part into the next part: that continuity is what paid off, in both accuracy and reproducibility.

Each row carries the question, the old answer, the new answer and a summary of the difference — down to the PMID, the year, and the exact passage in the source. A granularity you can use for training and for fact-checking alike.

• DPO (learning to prefer the better answer). The revision pairs drop straight in as material for preferring the newer recommendation.
• Medical education. Put old and new side by side to show what changed.
• Evaluating temporal understanding. Measure whether a model can tell the editions apart.
• Time-aware retrieval. Use the old edition as a tempting wrong answer to sharpen the search model.

CC BY 4.0. With attribution, it's free for commercial use and modification. The content is not a redistribution of the guideline text — it's differences and Q&A grounded in the retrieved passages. Always check against the original source before any clinical use.