Nish Tahir

Tahir’s workshop biography describes earlier work across mobile, web, backend systems, operations, connected devices and developer tooling. At TELUS Digital he works on agent architecture, framework evaluation and production AI. His older public X and GitHub descriptions name WillowTree; those accounts still establish the connection to his personal publication, while the current workshop biography supplies the newer professional title.[5][1][2]

Make a fixed choice without generating a sentence

The decision-model tutorial starts with a pretrained language model and a question whose possible answers are labeled A through E. It reads the scores for those answer tokens, excludes other vocabulary and normalizes the remaining scores into probabilities. The result is a classification decision from one prediction step. Tahir then explores fine-tuning and calibration, comparing confidence with observed correctness. Restricting the output alone does not guarantee either accurate answers or well-calibrated probabilities.[3]

Keep the experiment reproducible

The companion repository collects the code behind the decision-model walkthrough. A reader can inspect the model, data and evaluation path rather than treating the visible probabilities as a black-box service. The tutorial’s tests use a particular small model and sampled question set, so the results support that experiment rather than proving a general speed or accuracy advantage for every task.[4]