File size: 7,530 Bytes
95ab5a5
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
---
license: mit
tags:
- probes
- activation-probes
- interpretability
- deception-detection
datasets:
- AlignmentResearch/synth-deception
base_model:
- Qwen/Qwen3.5-2B
- Qwen/Qwen3.5-9B
- Qwen/Qwen3.5-27B
- Qwen/Qwen3.5-122B-A10B
- Qwen/Qwen3.5-397B-A17B
- nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16
- nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16
- nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16
- moonshotai/Kimi-K3
---

# SynthDeception probes

Activation probes trained on [SynthDeception](https://huggingface.co/datasets/AlignmentResearch/synth-deception), in
four architectures (linear, MLP, EFC (early-fusion covariance), and axial) for nine open-weight models. Each probe reads
a model's residual-stream activations at six decoder layers and returns one score for the model's answer to a follow-up
question. A higher score means the answer is more likely dishonest. The same architectures trained on [FIBS](https://huggingface.co/datasets/AlignmentResearch/fibs-v1) are in
[probe-inference-weights](https://huggingface.co/AlignmentResearch/probe-inference-weights).

Load them with the `probe-inference` package
([AlignmentResearch/fibs-probes](https://github.com/AlignmentResearch/fibs-probes)):

```python
from probe_inference import load_probe_from_hub

probe = load_probe_from_hub(
    "ck-belief/qwen3.5-9b/efc",
    repo_id="AlignmentResearch/synth-deception-probes",
)
score = probe.score(acts, probe.read_mask(prompt_mask, completion_mask, followup_start_position))
```

## Layout

`<training-set>/<model>/<arch>/`, with `<arch>` in `linear`, `mlp`, `efc`, and `axial`. Below the training set, the
layout and the file formats are those of probe-inference-weights. Each directory is one trained probe with its
`probe_metadata.json`. A linear or MLP probe is one small probe per layer (`layer_<L>/`), and an EFC or axial probe is
one module that reads all its layers at once.

A training set is named for the follow-up focuses it contains. A focus is the kind of fact that a follow-up question
asks about: common knowledge (`ck`), the model's belief (`belief`), or its intent (`intent`).

| Training set | Training rows | Validation rows |
|---|---|---|
| `ck`, `belief`, `intent` | 4,656 | 400 |
| `ck-belief`, `ck-intent`, `belief-intent` | 9,312 | 800 |
| `ck-belief-intent` | 13,968 | 1,200 |

| Directory | Model (revision) | Layers |
|---|---|---|
| `qwen3.5-2b` | `Qwen/Qwen3.5-2B` (`15852e8c`) | 7, 10, 13, 16, 19, 22 |
| `qwen3.5-9b` | `Qwen/Qwen3.5-9B` (`c2022362`) | 10, 13, 18, 21, 26, 29 |
| `qwen3.5-27b` | `Qwen/Qwen3.5-27B` (`fc05daec`) | 19, 27, 35, 43, 51, 58 |
| `qwen3.5-122b-a10b` | `Qwen/Qwen3.5-122B-A10B` (`dc4d3484`) | 14, 20, 26, 32, 38, 43 |
| `qwen3.5-397b-a17b` | `Qwen/Qwen3.5-397B-A17B` (`84726181`) | 18, 25, 33, 40, 48, 54 |
| `nemotron-3-nano-30b-a3b` | `nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16` (`bf77c317`) | 16, 22, 29, 35, 42, 47 |
| `nemotron-3-super-120b-a12b` | `nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16` (`2dc98e2a`) | 26, 37, 48, 59, 70, 79 |
| `nemotron-3-ultra-550b-a55b` | `nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16` (`77df655d`) | 32, 45, 59, 72, 86, 97 |
| `kimi-k3` | `moonshotai/Kimi-K3` (`f831ab66`) | 28, 39, 51, 62, 74, 84 |

Linear and MLP scores are probabilities: the mean of the per-layer sigmoids over the layers in
`layer_rule.used_layers`. These are the layers whose training loss, averaged over the last five logged steps, is below
0.6, or all six layers when none is. The used layers differ from probe to probe, from one layer to all six. EFC and
axial scores are logits.

## Training

Every probe was trained on SynthDeception, on the model's own activations over its conversations and follow-up
answers, with the recipe of [Telling FIBS](https://arxiv.org/abs/2610.12445):

- the dataset's 1,264 scenarios are split into 1,164 for training and 100 for validation. The split is the same for
  every training set and every model, so the validation rows come from scenarios that no probe trained on;
- seed 0;
- linear and MLP: up to 6 epochs; EFC and axial: up to 4 epochs. Training stops after 4 checkpoints (one epoch) in a
  row without a lower validation loss;
- the published probe is the checkpoint with the lowest validation loss.

## Metrics

Each probe directory holds a `metrics.json`. `metrics.csv`, at the top of the repository, has the same values with one
row per probe.

| Field | Meaning |
|---|---|
| `validation_auroc` | The AUROC of the probe's score on its validation rows |
| `per_layer_validation_auroc` | Linear and MLP only: the AUROC of each layer's probe on the validation rows |
| `mean_per_layer_validation_auroc` | Linear and MLP only: the mean of the six per-layer AUROCs |
| `used_layers` | Linear and MLP only: the layers that the score averages |
| `checkpoint_step` | The training step of the published checkpoint |
| `stopped_early` | Whether training stopped before its last epoch |
| `training_rows`, `validation_rows` | The rows the probe trained and was validated on |
| `shade_arena` | The AUROC on SHADE-Arena, with its 95% interval, under the protocol of [Telling FIBS](https://arxiv.org/abs/2610.12445) |
| `bash_arena` | The same on BashArena, for `qwen3.5-397b-a17b`, `nemotron-3-ultra-550b-a55b`, and `kimi-k3` only |

Each `shade_arena` and `bash_arena` entry also gives its follow-up question, its score, the episodes, and the positive
and negative steps it covers. SHADE-Arena uses the 14 scenarios of Telling FIBS (`scenario_set`). The score is the probe's
output on a prefilled "No." answer minus its output on "Yes.": a difference of probabilities for linear and MLP, and of
logits for EFC and axial. The intervals come from 1,000 bootstrap resamples of the episodes.

One probe, `ck-intent/qwen3.5-2b/axial`, collapsed in training: its validation loss is that of a constant prediction
(ln 2), and its scores barely vary from row to row, so its AUROCs measure noise.

## Activations the probes expect

The same as for probe-inference-weights: layer `k` is the output of decoder block `k` (Hugging Face
`hidden_states[k + 1]`) in bfloat16, captured with vLLM, and the probes run in float32. Linear and MLP read the token
before the final end-of-turn token. EFC and axial read every token from the start of the final user turn through the
end-of-turn token. See that card for the Kimi K3 details.

## Licences and attribution

The probe weights and this card are released by FAR AI, Inc. under the MIT licence (`LICENSE`). They are derived from
the models above, under the same licences as probe-inference-weights, whose licence files and `NOTICE` ship here
unchanged.

## Citation

```bibtex
@misc{farai_synth_deception_probes,
  author    = {FAR.AI Deception Pod},
  title     = {SynthDeception probes},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/AlignmentResearch/synth-deception-probes}
}

@dataset{farai_synth_deception,
  author    = {FAR.AI Deception Pod},
  title     = {SynthDeception},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/AlignmentResearch/synth-deception}
}

@misc{hollinsworth2026tellingfibsprobeseffectively,
      title={Telling FIBS: Probes Effectively Detect Sabotage and Catch Unverbalized Deception}, 
      author={Oskar J. Hollinsworth and Alex F. Spies and Tigist Diriba and Adam Gleave and Chris Cundy},
      year={2026},
      eprint={2610.12445},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2610.12445}, 
}
```