G4-MeroMero-31B-uncensored-heretic Slashes 85% Refusals For Creators
G4-MeroMero-31B-uncensored-heretic is a newly released language model that strips away almost all refusal behaviors. It builds on a fine-tuned version of Google’s Gemma 4 31B designed for storytelling and roleplay. Using abliteration, it rejects requests only 15% of the time, compared to 99% for the original.
Independent developer llmfan46 created this uncensored edition by applying abliteration to the Stardom Mero Mero model. The goal was to ease open-ended creative work by removing the moralizing pushback common in stock AI systems. He hosts over 70 free models and is asking for community help to cover storage costs.
Refusal rate cut by 85% with minimal knowledge loss
Key performance metrics
- Only 15 out of 100 prompts refused.
- 86.83% MMLU score, just 0.19% below original.
- KL divergence of 0.0100 confirms quality retention.
- Abliteration targets layers 28 through 49.
- Preserve good behavior weight set to 0.56.
- Built on Stardom Mero Mero creative finetune.
- Works in both thinking and non-thinking modes.
- Available in GGUF and NVFP4 quantizations.
Writers, roleplayers, and creative professionals get an AI that no longer lectures or refuses imaginative prompts. Privacy-focused users running local hardware can fit the 31B model on 24GB GPUs via 4-bit quantization. Content experimenters benefit from a capable assistant with minimal knowledge trade-off.
Storage crunch puts future uncensored models on hold
llmfan46 says his Hugging Face free storage is exhausted, blocking new model uploads without donations. He maintains more than 70 free models as an unpaid independent contributor. The abliteration method used careful parameters—a neighbor count of 10 and overcorrect weight of 0.9726—to suppress refusals while preserving accuracy.
"This is a decensored version of zerofata/G4-MeroMero-31B, made using Heretic v1.2.0 with the Arbitrary-Rank Ablation (ARA) method " — Source: Hugging Face
Project Links
