Gemma4-Gutenberg-31B

google/gemma-4-31B-it finetuned for literary, novelistic prose — a Gemma 4 entry in the Gutenberg series.

This pushes the (already strong) base toward a literary-fiction register: story and interiority over static description, controlled pacing over relentless adjective-stacking, and an active dispreference for "AI slop" phrasing.

Method

ORPO (Odds Ratio Preference Optimization) on the full schneewolflabs/Athanorlite-DPO (14,816 preference pairs) — a superset of the Gutenberg "Encore" recipe that bundles jondurbin/gutenberg-dpo-v0.1, nbeerbower/gutenberg2-dpo, gutenberg-moderne-dpo, human-writing-dpo, synthetic-fiction-dpo, Arkhaios-DPO, Purpura-DPO, Schule-DPO, sam-paech/gutenberg3, plus truthy / physical-reasoning / theory-of-mind balance sets.

Method	ORPO, β = 0.1
Adapter	LoRA r=64 (text decoder only), merged to full
LR	5e-5, cosine, 0.05 warmup
Epochs	1
Effective batch	32
Max length	2048
Optimizer	paged_adamw_8bit, bf16, grad-checkpointing
Hardware	1× NVIDIA GB10 (DGX Spark, 128 GB unified)
Trainer	Merlina (grimoire ORPO)

Training trajectory (clean convergence over ~3.5 days, 454 steps):

	start	end
eval/loss	2.4445	2.2052
reward_accuracy	0.175	0.9125
reward_margin	−0.076	+0.455

The reward_accuracy arc (0.18 → 0.50 → 0.91) reflects the model learning to prefer the literary chosen text while actively suppressing the rejected slop — the intended Gutenberg dynamic.

Notes

• Gemma 4 31B is a unified multimodal model; the vision/audio towers are left frozen and intact, so this remains a drop-in replacement for the base. Only the text decoder was tuned.
• This is a refinement of an already-capable writer, not a rescue — expect a consistent literary lean rather than a night-and-day transformation.

License

Apache-2.0 (matching the Gemma 4 base). Constituent training datasets carry their own licenses (see the Athanorlite-DPO card).