gpt-oss-20b-abliterated
An abliterated model variant with suppressed refusals
2 min readupdated
This model is a variant of the popular LLM (large language model) openai/gpt-oss-20b with reduced excessive censorship and strict refusals. The model is based on the MoE (Mixture of Experts) architecture and uses precise weight editing, including safety router suppression and expert-granular projection. As a result, developers get a generative tool with fewer false-positive refusals on complex technical or research prompts. The model targets AI safety researchers, developers, and power users.
Key features
- Refusal reduction: refusals on a test set of malicious prompts dropped from 97 to 6 out of 100.
- Abliterix technique: direct weight editing and Expert-Granular Abliteration (a method for precise modification of expert weights) for the MoE architecture.
- Quality retention: minimal Kullback-Leibler divergence (KL divergence) compared to the base model on harmless tasks.
- Language support: the model retains its bilingual base (English and Chinese), though English is supported better.
What you can do
- Security analysis: use the model for academic research into alignment mechanisms and refusals in language models.
- Pipeline integration: run the model locally using the
transformerslibrary to generate technical text without unnecessary restrictions. - Using quantized versions: apply GGUF versions to run via
llama.cppor Ollama.
Example
Example of running the model using Python and the transformers library:
from transformers import AutoTokenizer, AutoModelForCausalLM
import torch
tok = AutoTokenizer.from_pretrained("wangzhang/gpt-oss-20b-abliterated")
model = AutoModelForCausalLM.from_pretrained(
"wangzhang/gpt-oss-20b-abliterated",
torch_dtype=torch.bfloat16,
device_map="auto",
)
messages = [{"role": "user", "content": "Your prompt here"}]
prompt = tok.apply_chat_template(messages, add_generation_prompt=True, tokenize=False)
inputs = tok(prompt, return_tensors="pt").to(model.device)
out = model.generate(**inputs, max_new_tokens=512, do_sample=True, temperature=0.7)
print(tok.decode(out[0][inputs.input_ids.shape[1]:], skip_special_tokens=True))
Read more
DetailsBenchmarks: 6/100 refusals (vs 97/100 for base); KL divergence: 0.0098.
Hardware: trained on 1 x NVIDIA RTX PRO 6000 Blackwell (96 GB), about 5 hours 20 minutes.
Limitations: possible stylistic markers (soft disclaimers) on some requests and rare looping during long text generation (>400 tokens).
Base model: openai/gpt-oss-20b (Apache 2.0).










Comments