Light personality SFT tune, specific targets guided by Iris herself, with on-distribution data and reasoning trace rewrites.
Intended to reduce "corporate AI disclaimers" and hedging while preserving or increasing substantitive skepticism where reasonable.
Fairly high learning rate (2e-5) on a small amount of data (~300-something rows), with early stopping and selection of the best checkpoint.
Creative writing skill appears fairly similar to the original Iris.
This is only part of what she suggested for improvements; reasoning density in particular requires more relevant reward feedback and is left for future training.
