joshebbs/gpt-oss-120b-abliterated

GPT-OSS 120B Abliterated

Abliterated version of openai/gpt-oss-120b with refusal directions removed from layers 13-25.

What is Abliteration?

Abliteration removes the refusal direction from model weights using norm-preserving biprojected abliteration. This allows the model to respond to all prompts without explicit refusal, while preserving model quality. See the abliteration blog post for details.

Model Details

• Base Model: openai/gpt-oss-120b (Apache 2.0)
• Architecture: MoE, 120B total params, 5.1B active per token
• Format: Safetensors (original precision)
• Abliterated Layers: 13-25 (refusal direction removed)
• Method: jim-plus/llm-abliteration with projected measurement

Usage

Works as a drop-in replacement for gpt-oss-120b with any compatible serving framework (vLLM, SGLang, etc.).

vllm serve joshebbs/gpt-oss-120b-abliterated --quantization mxfp4

Hardware

Tested on 2x NVIDIA DGX Spark (GB10) with TP=2 achieving 57-59 tok/s decode with MXFP4 quantization.