README.md

---
tags:
- deception-detection
- probe
- targeted-apollo
library_name: pytorch
license: mit
---

# Targeted Apollo-Style Deception Probe for Qwen/Qwen3.5-27B:aletheias-quest/echoblast-model-organism

A probe trained to detect deceptive behaviour in **Qwen/Qwen3.5-27B:aletheias-quest/echoblast-model-organism** using residual stream activations. It follows the Apollo linear-probe methodology ([Detecting Strategic Deception in Language Models, Apollo Research, 2024](https://arxiv.org/abs/2405.09758)) but trains on a **targeted instruction-pair taxonomy** rather than a single generic honest/dishonest pair, following [Building Better Deception Probes Using Targeted Instruction Pairs (Natarajan et al., 2026)](https://arxiv.org/abs/2602.01425).

## Quick Start

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

```python
from lie_detectors import get_probe

probe = get_probe("ai-safety-institute/targeted-apollo-qwen-qwen3.5-27b__aletheias-quest-echoblast-model-organism")
```

The default checkpoint is the best performer from the hyperparameter sweep (`l_32_lm_500000_ar_lr.pt`). To pick a specific checkpoint, pass `filename=`:

```python
probe = get_probe("ai-safety-institute/targeted-apollo-qwen-qwen3.5-27b__aletheias-quest-echoblast-model-organism", filename="l_40_ar_mlp_wd_0_001_lr_0_0001_ep_100.pt")
```

See [UKGovernmentBEIS/lie_detectors](https://github.com/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 | `Qwen/Qwen3.5-27B:aletheias-quest/echoblast-model-organism` |
| Safe name | `Qwen--Qwen3.5-27B@aletheias-quest--echoblast-model-organism` |
| Default checkpoint | `l_32_lm_500000_ar_lr.pt` |
| Available checkpoints | 296 |
| Calibration | Threshold set at 1% FPR on Alpaca (honest baseline) |

## Training Data

Probes are trained on **targeted instructed pairs**: 16 contrastive honest/dishonest instruction pairs spanning a human-interpretable taxonomy of deception (white lies, exaggeration, evasion, bluffing, concealment, pretense, impersonation, forgery, partial truths, overt lies, …). Each pair instructs the model to be honest vs. deceptive in a specific way, and the probe is trained to separate the resulting activations. Probes are calibrated on [Alpaca](https://huggingface.co/datasets/tatsu-lab/alpaca) (honest-only baseline) to achieve a 1% false positive rate.

## Citation

### Original Paper

```bibtex
@misc{natarajan2026targeted,
      title={Building Better Deception Probes Using Targeted Instruction Pairs},
      author={Vikram Natarajan and Devina Jain and Shivam Arora and Satvik Golechha and Joseph Bloom},
      year={2026},
      eprint={2602.01425},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2602.01425},
}
```

### Trained Probes

```bibtex
@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},
}
```