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

LatentQA baseline adapter for Qwen3.5-9B

This repository contains the Qwen3.5-9B LoRA adapter used for the LatentQA comparison in the PRISM demo and evaluation. It is a baseline artifact, not a PRISM checkpoint. The adapter reads target-model activations and answers a free-form question about them.

The implementation is adapted from aypan17/latentqa at revision a2dcb6f. The published adapter omits duplicated embedding and language-head tensors; the PRISM demo attaches it to a separately downloaded Qwen3.5-9B model.

Files

File	SHA-256
adapter_config.json	a952a74b6be55979834a491a7da2f1ddfff0f9ddc153116b692c6b6ab4a220a6
adapter_model.safetensors	8600f3ba51e60e53de72ff044adee962dca3c179a9b8a1212809c236a7852cd2

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 latentqa
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 LatentQA implementation is Apache-2.0 and is attributed in the PRISM repository.

References

• Alexander Pan, Lijie Chen, and Jacob Steinhardt. LatentQA: Teaching LLMs to Decode Activations Into Natural Language.
• Gilad Gressel et al. PRISM: Recovering Instruction Sets from Language Model Activations. EMNLP 2026.