HuiHui abliterated model outperforms vanilla 3.6-35B-a3b on math and code

A custom quantization recipe applied to the HuiHui abliterated model demonstrates superior performance compared to the vanilla 3.6-35B-a3b variant in mathematics and coding tasks. The results suggest that removing refusal mechanisms allows the model to achieve greater accuracy and wisdom in these domains.

• Identical custom quantization was used for both HuiHui and Vanilla 3.6-35B-a3b models.
• Benchmarks were conducted using the oMLX suite in instruct mode without long reasoning chains.
• The abliterated model showed improved performance on math and coding benchmarks despite a small sample size.

The findings indicate that abliteration can enhance model capabilities in specific technical areas, with the HuiHui model available for download via the provided Hugging Face link.