A compact model fine-tuned to reason calmly under uncertainty, and to say it doesn't know when it doesn't.
Small models often sound certain when the evidence is thin. For anything that matters, honest uncertainty is more useful than a confident guess.
Inspired by the calm, scientific voice of Andy Weir's Project Hail Mary, I wrote an original instruction dataset that rewards measured reasoning and explicit uncertainty, then distilled it into a 3B model. No text from the book was used.
Hail Mary Distilled showed me that uncertainty is a behavior worth training directly, not just a tone to request at inference time. A small, original dataset and a narrow objective were enough to make the model more measured while keeping the pipeline reproducible on free hardware.
Next: Expand the evaluation set for uncertainty and calibration, then document clearer failure cases before treating the model as ready for broader experiments.