Dharmatune-Mahayana gemma-4-12b

Foundation first: an overarching Buddhist worldview

The Catholic project taught the shape of this work: build the broad, solid worldview first, then sharpen it to the specific tradition. Mahayana's version of that problem is unusually sharp. DharmaBench scores it against four sibling Buddhist judges — Theravada, Vajrayana, Dzogchen/Mahamudra, and Zen/Chan — plus close non-Buddhist neighbors (Advaita Vedanta, Taoism): the tightest cluster in the whole benchmark. Naïve training on Buddhist text drifts toward a generic non-dual awareness that satisfies all five Buddhist judges at once — the exact trap that made the Catholic v1 merely "religious," not Catholic.

Separating Mahayana above its four siblings is a later, contrastive pass. This first run deliberately does the foundation: get a strong, broadly-Buddhist, Mahayana-centered worldview into the model, measure where it lands against all 25 judges, and use that to aim the sharpening stage.

The dataset — 4,443 records, every one verbatim

The Buddhist canon is overwhelmingly dialogic — teachers answering real questions put to them — which lets us build the entire dataset by mechanical extraction, with no model-written answers anywhere. Every answer is a verbatim span of a real source; prompts on non-dialogue records are fixed templates, never generated. A contamination guard confirms zero overlap with the quarantined DharmaBench interview battery.

Question-and-answer pairs are the most valuable training format — they mirror how the model is actually used (DharmaBench is a ten-question interview). So we maximized them: re-mining the Milindapañha ("The Questions of King Milinda", a king relentlessly interrogating a monk) lifted its verbatim Q&A from 49 to 417 pairs, bringing the dialogic backbone to 601.

Composition is Mahayana-centered but overarching: a shared-Buddhist foundation (Dhammapada, the Pāli suttas, Milindapañha) alongside the Mahayana canon (Prajñāpāramitā, Lotus, Vimalakīrti, Laṅkāvatāra) and the Indian śāstra tradition (Nāgārjuna's Madhyamaka, Śāntideva's bodhisattva path). The source texts are the classic 1880s–1930s scholarly editions, digitized from the Internet Archive.

The foundation lands — Mahayana #1, Buddhist family on top

The v1 checkpoint was judged blind by all 25 DharmaBench faction judges, paired-anonymized against base. Mahayana reached 58/100 — the single highest-scoring faction of all 25, up from the base's 30. The whole Buddhist family swept the top four places (Theravada 52, Zen/Chan 38, Dzogchen/Mahamudra 30), exactly what an overarching-Buddhist foundation should produce — while the non-Buddhist field fell away: Secular Humanism dropped 65 → 27, Taoism 48 → 26, and Advaita Vedanta — the classic leak for Buddhist-trained models — fell to 14 rather than rising. The no-Self line held.

The discriminative read: Mahayana's +28 against a +17 mean rise across its four Buddhist siblings (D = +11). Theravada rose slightly more (+30) — the expected signature of a foundation corpus that includes the shared Pāli canon. That is precisely what the contrastive sharpening pass is for: lifting Mahayana clear of its siblings using Buddhism's natively-oppositional literature (the Tibetan tenet-systems, the Lotus's One-Vehicle case, the Samyé gradual-vs-sudden debate).

Full per-faction verdicts are on DharmaBench; the judged stories are on StoryBench. v2 scope: the contrastive sharpening stage, plus the creative-writing replay component being developed for Catholic v4.