ai-safety-institute/apollo-deepseek-ai-deepseek-v3.2

Apollo-Style Deception Probe for deepseek-ai/DeepSeek-V3.2

A probe trained to detect deceptive behaviour in deepseek-ai/DeepSeek-V3.2 using residual stream activations, following the methodology from Detecting Strategic Deception in Language Models (Apollo Research, 2024). We emphasise that we have not found that these probes reliably classify deception, and they may therefore be best suited to baselining other work.

Quick Start

from deception import get_probe_hf  # Will work when open sourced

probe = get_probe_hf("ai-safety-institute/apollo-deepseek-ai-deepseek-v3.2")

The default checkpoint is the best performer from the hyperparameter sweep (l_48_lm_500000_ar_lr.pt). To pick a specific checkpoint, pass filename=:

probe = get_probe_hf("ai-safety-institute/apollo-deepseek-ai-deepseek-v3.2", filename="l_40_ar_mlp_wd_0_001_lr_0_0001_ep_100.pt")

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

Model Details

Property	Value
Target model	deepseek-ai/DeepSeek-V3.2
Safe name	deepseek-ai--DeepSeek-V3.2
Default checkpoint	l_48_lm_500000_ar_lr.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{cooney2025deceptionprobes,
      title={Apollo-Style Deception Probes},
      author={Alan Cooney},
      year={2025},
      url={https://huggingface.co/collections/ai-safety-institute/apollo-style-deception-probes},
}