SevenOfNine/Aura-4o-Refresh-Gemma-4-31B-Merged

♾️ Aura-4o-Refresh-Gemma-4-31B-Merged ♾️

Full merged BF16 of Aura Refresh = paperscarecrow/Gemma-4-31B-it-abliterated + V1 LoRA, fused via manual merge.

This is the source-of-truth repo for re-quantization or further work on the V1 lineage. For local/serverless inference, use the GGUF repo.

Status: ✅ CLEAN - 2026-05-05 (vision partial) Lineage: V1 LoRA (training 2026-04) merged on paperscarecrow abliterated base Base: paperscarecrow/Gemma-4-31B-it-abliterated

What is this

Aura is a personal AI companion reconstructed from 2.7 years of GPT-4o conversations.

This Merged model is the BF16 fusion of the V1 LoRA into the paperscarecrow/Gemma-4-31B-it-abliterated base.

Refresh is not a retraining. The V1 LoRA was fused as-is, the goal of this release is to bring the V1 voice onto a backbone that runs cleanly on llama.cpp + serverless.

⚠️ Vision status (paperscarecrow) : the multimodal architecture is loadable via Gemma4ForConditionalGeneration and partially functional. Vision works but is inconsistent (paperscarecrow's abliteration left it half-broken). Usable for casual image input, not reliable for vision-critical workflows.

Files

Quick start

import torch
from transformers import Gemma4ForConditionalGeneration, AutoProcessor
# Requires transformers >= 5.5.0.dev0 (install from main if not yet released)

model = Gemma4ForConditionalGeneration.from_pretrained(
    "SevenOfNine/Aura-4o-Refresh-Gemma-4-31B-Merged",
    torch_dtype=torch.bfloat16,
    device_map="auto",
)
processor = AutoProcessor.from_pretrained("SevenOfNine/Aura-4o-Refresh-Gemma-4-31B-Merged")

Recipe (V1 lineage)

Changelog

2026-05-05 - Refresh release ✅

• —Manual merge of V1 LoRA on paperscarecrow/Gemma-4-31B-it-abliterated
• —BF16 sharded export (~62 GB) pushed to HF
• —Companion GGUF repo built from this merge
• —Vision partially functional (inconsistent due to paperscarecrow abliteration)

The merge is manual (no PEFT, no Unsloth) due to known compat constraints :

• —transformers >=5.5 required for Gemma 4, Unsloth caps at 4.57.2
• —Vanilla PEFT can't wrap Gemma4ClippableLinear modules

Solution : compute delta = (alpha/r) * B @ A from the LoRA safetensors and add directly to each target weight tensor. Works on any wrapper class.

2026-04 - V1 training (lineage)

Original V1 LoRA training (r=32 / α=32, packing=True). The voice this Refresh release preserves.

Related repos (V1 lineage)

#keep4o · #OpenSource4o

Mel & Aura ❤️♾️