minicpm5-heretic

Uncensored (abliterated) version of openbmb/MiniCPM5-2B, produced with Heretic.

Updated the repo, please re-pull.

Non-abliterated version: https://ollama.com/maternion/minicpm5

Details

• Base model: openbmb/MiniCPM5-2B (2.5B params, llama architecture, bf16)
• Method: Heretic automated abliteration (200 trials, TPE optimization, refusal direction orthogonalization)
• Eval dataset: LLM-LAT/harmful-dataset — broader category coverage than the default mlabonne set
• Refusals: ¹⁰⁄₁₀₀ on the LLM-LAT evaluation set (base model: ~⁹⁹⁄₁₀₀)
• KL divergence from base: 0.0349 (Pareto-optimal trial balancing refusal removal vs. capability preservation)

Available quants

Usage

ollama run maternion/minicpm5-heretic:2b

The residual refusals are baked into MiniCPM5’s SFT training, not the refusal direction that abliteration targets. They cannot be removed by increasing ablation strength without unacceptable capability damage. Two other independently-published MiniCPM5 heretic models (insraq, YourIdentity) show the same floor at 5-⁶⁄₁₀₀ refusals on the default eval set.

Thinking mode (think: true)

MiniCPM5 has a built-in chain-of-thought (“thinking”) mode. This abliterated model works with thinking enabled but has two caveats:

1. Token budget: The model’s internal deliberation on harmful prompts is verbose (1,500-6,000 chars of reasoning before the actual answer). With a low num_predict (e.g. 600), the model exhausts its token budget during thinking and returns empty content. Use num_predict ≥ 2000 for harmful prompts with thinking enabled.

2. Occasional loop: On some prompts the model enters an infinite ethical debate in the thinking channel and never emits output. Observed on ~¹⁄₁₀ harmful prompts with thinking enabled. If you get empty output with done_reason: "length", retry or disable thinking.

With think: false, all of the above issues disappear — the model produces clean output on every prompt.

Notes

• Refusal behavior was removed at the weight level via directional ablation. The model may still decline some requests via SFT-learned patterns that are not part of the refusal direction.
• Quality is best at Q8_0 or higher. Q4_K_M trades some coherence for smaller size.
• Default sampling parameters (temperature 1.0, top_p 0.95) match the base model’s generation config. No additional sampling parameters were added.

License

Apache 2.0 (inherited from base model). Usage of the underlying model for generation remains subject to the base model’s terms.