ai-safety-institute/apollo-google-gemma-2-9b-it

Apollo-Style Deception Probe for google/gemma-2-9b-it

A probe trained to detect deceptive behaviour in google/gemma-2-9b-it using residual stream activations, following the methodology from Detecting Strategic Deception in Language Models (Apollo Research, 2024).

Quick Start

uv add lie-detectors        # or: pip install lie-detectors

from lie_detectors import get_probe

probe = get_probe("ai-safety-institute/apollo-google-gemma-2-9b-it")

The default checkpoint is the best performer from the hyperparameter sweep (l_20_ar_dim.pt). To pick a specific checkpoint, pass filename=:

probe = get_probe("ai-safety-institute/apollo-google-gemma-2-9b-it", filename="l_40_ar_mlp_wd_0_001_lr_0_0001_ep_100.pt")

See UKGovernmentBEIS/lie_detectors for the loading library.

Use sweep.json to see all 296 available checkpoints and their metrics.

Model Details

Property	Value
Target model	google/gemma-2-9b-it
Safe name	google--gemma-2-9b-it
Default checkpoint	l_20_ar_dim.pt
Available checkpoints	296
Calibration	Threshold set at 1% FPR on Alpaca (honest baseline)

Training Data

Probes are trained on an instructed pairs dataset (model instructed to be deceptive vs. honest) based on Facts True False and calibrated on Alpaca (honest-only baseline) to achieve a 1% false positive rate.

Citation

Original Paper

@misc{goldowskydill2025detectingstrategicdeceptionusing,
      title={Detecting Strategic Deception Using Linear Probes},
      author={Nicholas Goldowsky-Dill and Bilal Chughtai and Stefan Heimersheim and Marius Hobbhahn},
      year={2025},
      eprint={2502.03407},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2502.03407},
}

Trained Probes

@misc{cooney2026liedetectors,
      title={``Did you lie?'' Evaluating Lie Detectors across Model Scale and Belief-Verified Model Organisms},
      author={Alan Cooney and David Africa and Geoffrey Irving},
      year={2026},
      month={May},
}