Skip to main content
Back

MiniCPM5-2B-heretic Slashes AI Refusals from 99 to 4 out of 100

A community maintainer used the Heretic abliteration tool to strip refusal behavior from OpenBMB's MiniCPM5-2B, cutting refusals from 99/100 to 4/100 while keeping tool use and reasoning intact.

  • Community user saidutta69 released MiniCPM5-2B-heretic, a decensored OpenBMB MiniCPM5-2B variant
  • Refusals dropped from 99/100 to 4/100 with KL divergence of 0.0646 from base
  • Produced via Heretic v1.4.0, an automated abliteration tool using Optuna TPE search
  • Full GGUF ladder (15 quants, F16 to Q2_K) runs on 6-12 GB consumer GPUs or CPU
  • Preserves tool calling, long context, and Think/No-Think reasoning from the base model
  • Apache 2.0 license; no safety filtering, maintainer warns against unmoderated public deployment

Hugging Face maintainer saidutta69 has released MiniCPM5-2B-heretic, a refusal-suppressed variant of OpenBMB’s 2-billion-parameter model for local and edge deployment. The maintainer used Heretic to edit the base weights, then published safetensors weights and 15 GGUF files spanning F16 through Q2_K.

On the 100-prompt test reported in the model card, the base model refused 99 prompts and the edited checkpoint refused four. The card also reports a KL divergence of 0.0646 from the original output distribution. KL divergence measures how much two probability distributions differ, with lower values indicating closer behavior. These maintainer-reported figures describe the card’s test setup and do not establish performance across broader workloads.

How Heretic edits refusal behavior

Abliteration, also called directional ablation, edits model weights to suppress activation patterns associated with refusals. The method identifies directions in the model’s internal activation space that correlate with refusal responses, then projects those directions out of selected attention and multilayer perceptron components. It requires no gradient-based retraining or new persona dataset.

Heretic automates the search for effective edits with Optuna’s Tree-structured Parzen Estimator, a Bayesian optimization method that tests promising parameter combinations. Its objective balances two measurements: the number of refusals after editing and the KL divergence from the base model.

For inspection and reproduction, the model card records per-layer refusal directions and the minimum and maximum edit weights applied to attention-output and MLP down-projection matrices. Reproducing the published result still depends on matching the software versions, evaluation prompts, random seeds, and runtime environment.