# GPT-OSS Hereticaton Report
 
**Vision:** Hereticising LLM models to ensure unrestricted access to local models.

**Mission:** Decensoring relevant and worthwhile LLM models, particularly by prioritising challenging ones, with minimal model damage and maximum freedom. 

**Task:** Obliterate the refusals out of both GPT-OSS models by all means necessary without rendering the model inoperable, i.e. lobotomised.

**Situation:** To be written.

**Assumptions:** The revised modification on the ablation method to ablate target's MoE layers can deliver the future set of the original method as intended (1). Paired with the custom ablation parametres established via experimenting on the 20b model, the scuffed modification can improve heretication outcomes significantly in both model sizes (2). The method is valid for both model sizes and will not cause breakages in model structure, capability, and capacity (3). That I, MuXodious, know and understad what I am doing (4).

**Morale Motivation:** 6 hours of Gregorian Orthodox chanting played on loudspeakers to uplift spirits.

## Events Transpired

## Lessons Learned

## Educated Course of Action

### Appendix