GPT-OSS 120B Model Weight Modification Expert

I need an experienced AI engineer to develop a complete Python solution for modifying the actual weights of GPT-OSS 120B, including its built-in refusal mechanisms, not just system prompts. My setup: * Vast.ai server with NVIDIA GB10 * GPT-OSS 120B in MXFP4 format * 14 model weight files * Independent backup already completed and verified I need working code, compatibility testing, and clear instructions. Do not delete or overwrite the original model files. Please confirm your experience with LLM weight editing, MXFP4 quantization, and GPT-OSS architecture before starting.