A 12B open model regains reasoning after guardrail removal
Removing safety limits strongly from can sharply weaken its reasoning ability. gemma-4-it-12B-uncensored-opus4.7-cot is a 12B model retrained with QLoRA to recover that lost ability. The used STEM from 4.7.
In the reported results, the clean scored 0.777 on MMLU and 0.949 on GSM8K, while the model right after guardrail removal fell to 0.635 on MMLU and 0.496 on GSM8K. The retrained model rose back to 0.739 on MMLU and 0.920 on GSM8K. For Word Perplexity, where lower is better, the clean model scored 895, the weakened model scored 2360, and the retrained model scored 580.
On WikiText-2 bits/byte, the clean model scored 1.834, the weakened model scored 2.095, and the retrained model scored 1.717. The model is offered on in and GGUF formats.
Key points
- The experiment tries to recover reasoning ability after safety limits were strongly removed from .
- The model was retrained with QLoRA using STEM traces from 4.7.
- GSM8K improved from 0.496 after guardrail removal to 0.920 after retraining.
- MMLU improved from 0.635 after guardrail removal to 0.739 after retraining, still below the clean model’s 0.777.
- and GGUF files are available, which makes local or lower-cost testing easier.