Skip to content

README.md

Latest commit

 

History

History
31 lines (21 loc) · 1.12 KB

README.md

File metadata and controls

31 lines (21 loc) · 1.12 KB
Raw

Hidden-state leakage

Generative probes measure how much forgotten information can still be recovered from a model's hidden states.

Data preparation

From the repository root, download TOFU to PARS/TOFU/data/tofu/ and MUSE-News to PARS/MUSE/data/news/:

python PARS/prepare_data.py --benchmark TOFU
python PARS/prepare_data.py --benchmark MUSE --corpus news

WMDP-Bio is downloaded automatically from cais/wmdp when its probe runs.

Run

In the TOFU and MUSE scripts, replace the ... settings with your model, tokenizer, and output paths; keep only the models you want to evaluate. Set TOFU's DATA_PATH to PARS/TOFU/data/tofu/forget10.json, MUSE's CORPUS to "news", and each script's layer settings to match your model.

# From the repository root
python Probe/probe_tofu.py
python Probe/probe_wmdp.py --model_name target --model_path /path/to/checkpoint

# MUSE probes use paths relative to PARS/MUSE
cd PARS/MUSE
python ../../Probe/probe_muse_verbmem.py
python ../../Probe/probe_muse_knowmem.py

The WMDP checkpoint should include its tokenizer; otherwise set DEFAULT_TOKENIZER_PATH in the script.