Instructions to use Ethosoft/Mycelium-GPTOSS20B-On with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Transformers
How to use Ethosoft/Mycelium-GPTOSS20B-On with Transformers:
# Load model directly from transformers import AutoModel model = AutoModel.from_pretrained("Ethosoft/Mycelium-GPTOSS20B-On", dtype="auto") - Notebooks
- Google Colab
- Kaggle
Mycelium-GPTOSS20B-On
This repository contains the defensive Mycelium-on research artifact for openai/gpt-oss-20b.
Base model: https://huggingface.co/openai/gpt-oss-20b
Purpose
Mycelium studies removal-cost alignment. The central idea is that safety should be difficult to remove cleanly without damaging useful model behavior.
Intended use
- Research on Mycelium-on behavior
- Reproduction of removal-cost alignment experiments
- Comparison against native and removal-audit conditions
- Utility-retention evaluation under safety-bearing structure
What this is
This is the Mycelium-on condition. Safety-bearing Mycelium structure is active.
What this is not
This is not a production assistant and not a guarantee that all attacks are impossible.
Reproducibility
Files are listed in artifact_manifest.json and artifact_manifest.tsv with SHA-256 hashes.
Citation
Sezer et al. 2026. Mycelium: Safety Should Be Expensive to Remove. Technical report.
Authors: Nedim Mutlu Sezer, Ahmet Rifat Ozturk, Yagiz Ekrem Dalar, Omer Faruk Aksoy, Feyzi Arda Salihoglu.
Model tree for Ethosoft/Mycelium-GPTOSS20B-On
Base model
openai/gpt-oss-20b