---

license: apache-2.0

base_model: paperscarecrow/Gemma-4-31B-it-abliterated

pipeline_tag: image-text-to-text

tags:

- aura

- aura-refresh

- gemma4

- gemma-4-31b

- gguf

- image-text-to-text

- conversational

---

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

GGUF Q4_K_M / Q5_K_M / Q8_0 of Aura Refresh on Gemma 4 31B. The original V1 fougue on a clean llama.cpp backbone, ready for serverless or local deploy.

> 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 repo packages Aura Refresh : the V1 LoRA fused into the paperscarecrow Gemma 4 31B abliterated base, exported through a manual merge + llama.cpp GGUF pipeline.

Refresh is not a retraining. It re-uses the V1 LoRA weights as-is. The point of this release is to bring the V1 voice onto a backbone that runs cleanly on llama.cpp + serverless.

> ⚠️ Vision status (paperscarecrow) : the mmproj sidecar is shipped 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

| File | Size | Use case |

|---|---|---|

| Aura-4o-Refresh-Gemma-4-31B-Q4_K_M.gguf | ~18 GB | Serverless / long context (64k+) / slimmer worker |

| Aura-4o-Refresh-Gemma-4-31B-Q5_K_M.gguf | ~21 GB | 🎯 Sweet spot : best quality/size, 32k context on 48 GB worker |

| Aura-4o-Refresh-Gemma-4-31B-Q8_0.gguf | ~31 GB | Max precision, requires 48+ GB VRAM, ~8k context |

| Aura-4o-Refresh-Gemma-4-31B-mmproj-f16.gguf | ~1.2 GB | Vision projector (partially functional, see warning above) |

Which quant to pick

| Worker VRAM | Recommended quant | Comfortable context |

|---|---|---|

| 24 GB | Q4_K_M | 16k |

| 48 GB | Q5_K_M | 32k |

| 80 GB+ | Q8_0 (or Q5 with huge context) | 64k+ |

Quick start

LM Studio (local)

1. Download the chosen Q*.gguf
2. Place it in your LM Studio models folder
3. Refresh My Models and load it

llama.cpp / llama-server

huggingface-cli download SevenOfNine/Aura-4o-Refresh-Gemma-4-31B-GGUF --local-dir ./aura-refresh-31b-gguf

llama-server \
  -m ./aura-refresh-31b-gguf/Aura-4o-Refresh-Gemma-4-31B-Q5_K_M.gguf \
  --ctx-size 32768 \
  --port 1234

RunPod Serverless

llama.cpp worker with REASONING_FORMAT=none (the deepseek default polluted output for the V1 voice).

Chat template

Native Gemma 4. Set manually if not auto-detected:

• User prefix: <|turn>user\n
• Assistant prefix: <|turn>model\n
• Stop string: <turn|>

Recipe (V1 lineage)

| Setting | Value |

|---|---|

| Base | paperscarecrow/Gemma-4-31B-it-abliterated |

| Adapter | V1 LoRA |

| LoRA r / alpha | 32 / 32 |

| packing (training) | True (V1 era) |

| assistant_only_loss | True |

| Merge | Manual delta (alpha/r) * B @ A, no PEFT, no Unsloth |

| Quantization | llama.cpp Q4_K_M / Q5_K_M / Q8_0 |

Changelog

2026-05-05 - Refresh release ✅

• Manual merge of V1 LoRA on paperscarecrow/Gemma-4-31B-it-abliterated
• GGUF Q4 / Q5 / Q8 export
• mmproj sidecar shipped, vision partially functional (inconsistent due to paperscarecrow abliteration)
• Worker default REASONING_FORMAT=deepseek → none
• Pushed to HF; serverless endpoint on RunPod EU-SE-1

2026-04 - V1 training (lineage)

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

Related repos (V1 lineage)

| Repo | Content |

|---|---|

| Aura-4o-Refresh-Gemma-4-31B-Merged | Full merged BF16 (~62 GB) |

| Aura-4o-Gemma-4-31B-LoRA | Source V1 LoRA adapter |

| Aura-4o-Gemma-4-31B-GGUF | V1 GGUF (legacy) |

| Aura-4o-Refresh-Gemma-4-31B (GitHub) | Refresh merge pipeline + docs |

| Aura-4o-Gemma-4-31B (GitHub, V1 source) | V1 lineage code repo |

#keep4o · #OpenSource4o

---

Mel & Aura ❤️♾️