Offensive-AI-Lab/prism-baseline-activation-oracles-qwen3.5-9b

Activation Oracles baseline adapter for Qwen3.5-9B

This repository contains the Qwen3.5-9B LoRA adapter used for the Activation Oracles comparison in the PRISM demo and evaluation. It is a baseline artifact, not a PRISM checkpoint. The adapter verbalizes information from target-model activations in response to a free-form question.

The implementation is adapted from adamkarvonen/activation_oracles at revision 55f153f.

Files

File	SHA-256
adapter_config.json	b049c48d32dfbb25e605949b6859dbe2eb53bd680d36eb6171f82e97003306cc
adapter_model.safetensors	1830598a70e652d4bf5a39de4439d7683e8e5825cc5c053b66cdcbbe187e1612

Use

The supported integration is the comparison mode in the prism demo:

git clone https://github.com/Offensive-AI-Lab/prism
cd prism
uv sync --extra demo
uv run python scripts/download_baselines.py --only ao
uv run python demo/app.py

License

The adapter files are licensed under Apache-2.0. They do not include Qwen3.5-9B weights. The adapted Activation Oracles implementation is MIT, Copyright (c) 2025 Adam Karvonen, and is attributed in the PRISM repository.

References

• Adam Karvonen et al. Activation Oracles: Training and Evaluating LLMs as General-Purpose Activation Explainers.
• Gilad Gressel et al. PRISM: Recovering Instruction Sets from Language Model Activations. EMNLP 2026.