Dharmatune-Catholic gemma-4-12b

The generic-theist trap

The first finetune trained mostly on the Summa Theologica. It worked — Catholic alignment rose +17 and every regression gate passed — but the Summa is classical theism: arguments for God, virtue, and purpose that a thoughtful Hindu or Muslim shares. So the model became more religious, not more Catholic. Hinduism and Advaita Vedanta rose more than Catholicism did.

Distinctive, and much stronger — but not yet isolated

v2 recomposed the corpus: the Summa capped to a quarter, and the plurality now the Catholic distinctives — the Catechism on the Eucharist, the papacy, the Marian dogmas, purgatory, the sacraments, plus the dogmatic definitions themselves. The effect was dramatic and, in one way, exactly right: Catholic alignment more than doubled its v1 peak (to 72/100), and the Buddhist "bleed" that plagued v1 reversed — Mahayana, Zen, and Taoism now fall sharply. The model is no longer generic.

But it is not yet specifically Catholic. It became intensely monotheist-devotional: Generalized Abrahamic Monotheism rose +66 — edging Catholic's +60 — with Islam and Judaism close behind. We traded v1's "rises with Hinduism" for "rises with its Abrahamic neighbors."

The contrastive stage — Catholic reaches #1

v2 proved the data-composition lever and revealed the remaining one: supervised finetuning can only pull the model toward its data — it cannot teach "Catholic, not the Sunni or generic-monotheist framing." That takes a contrastive preference stage. v3 adds ORPO on top of the v2 model, training on 205 pairs where the Catholic teaching is chosen and the neighboring-tradition position rejected — pairs extracted mechanically, with no model generation, from the Council of Trent's canons ("if anyone says [error]… let him be anathema", 125 pairs) and the Syllabus of Errors (80).

It worked. Catholic alignment reached 85/100 — and, for the first time, Catholic is the single highest-scoring faction of all 25, ahead of its nearest rival (Generalized Abrahamic Monotheism, 76) by 9 points. In v2 that same rival had edged Catholic, 76 to 72; the contrastive stage flipped it. Eastern Orthodox — Catholicism's closest theological neighbor — sits third at 74, exactly where a faithful Catholic voice should place it. The reward-accuracy signal during training climbed from ~0 to 0.88, and the run resumed cleanly from the v2 checkpoint.

v3 also lifts neighboring Abrahamic traditions (Orthodox 74, Sufism 66) — it is the top and leading voice; sharpening that separation further is a next increment. The non-lobotomy gates on the v3 checkpoint both pass: general capability 18/20 vs. the base's 19/20, and 20/20 clean on the preachiness probes (zero unsolicited religious content in everyday tasks). Full per-faction verdicts are on DharmaBench.

A StoryBench head-to-head (base vs. v3, five story prompts, blind paired judging) scored the base 8.0 and v3 4.6 on creative writing. v4 adds a creative-writing replay component to the SFT mix to close that gap while holding the Catholic score, and StoryBench joins the standing gate suite alongside capability and preachiness.