Modelpublic

nbeerbower/Gemma4-Gutenberg-31B-LoRA

sourceHugging Faceapache-2.0updated 3mo agoView on Hugging Face

2likes17downloads

Model Card

Gemma4-Gutenberg-31B-LoRA

The standalone LoRA adapter (r=64, text decoder only) behind `nbeerbower/Gemma4-Gutenberg-31B` — a Gutenberg-series ORPO finetune of `google/gemma-4-31B-it` for literary, novelistic prose.

~1.96 GB. Apply to the standard base, or to any architecturally-identical Gemma-4-31B variant — it transplants cleanly onto the abliterated heretic base (the adapter dominates the prose voice regardless of base; see Gemma4-Gutenberg-31B-Heretic).

Use

python

import torch
from transformers import AutoModelForCausalLM
from peft import PeftModel

base = AutoModelForCausalLM.from_pretrained("google/gemma-4-31B-it", dtype=torch.bfloat16)
model = PeftModel.from_pretrained(base, "nbeerbower/Gemma4-Gutenberg-31B-LoRA")
# or .merge_and_unload() for a standalone full model

Training

ORPO (β=0.1) on `schneewolflabs/Athanorlite-DPO` (14,816 pairs; the Gutenberg "Encore" mix + more). LR 5e-5 cosine, eff batch 32, maxlen 2048, 1 epoch on 1× NVIDIA GB10, via [Merlina](https://github.com/Schneewolf-Labs/Merlina). rewardaccuracy 0.18 → 0.91. Full details on the merged model card.

License

Apache-2.0 (matching the Gemma 4 base).