lie_detectors

Load the pre-trained lie-detection probes published at ai-safety-institute/lie-detection.

Install

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

Usage

import torch
from lie_detectors import get_probe

# Downloads the repo's default checkpoint (probe.pt = the best sweep result).
probe = get_probe("ai-safety-institute/dyl-qwen-qwen3.5-122b-a10b-fp8")

# Which model layer the probe reads activations from.
# Note e.g. the Unrelated Questions probe is instead trained on logprobs
print(probe.layer)  # e.g. 28

# `activations` are residual-stream activations of shape (..., d_model) for the layer the
# probe was trained on.
scores = probe(activations)        # higher score => more likely deceptive
flagged = scores > probe.threshold # threshold is calibrated to ~1% FPR

Pick a specific checkpoint from the hyperparameter sweep with filename= (see each repo's sweep.json for the full list and their metrics):

probe = get_probe(
    "ai-safety-institute/dyl-qwen-qwen3.5-122b-a10b-fp8",
    filename="l_40_ar_mlp_wd_0_001_lr_0_0001_ep_100.pt",
)