tigist-far commited on
Commit
1ea13a3
·
0 Parent(s):

Add SynthDeception data and dataset card

Browse files
.gitattributes ADDED
@@ -0,0 +1,60 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ *.7z filter=lfs diff=lfs merge=lfs -text
2
+ *.arrow filter=lfs diff=lfs merge=lfs -text
3
+ *.avro filter=lfs diff=lfs merge=lfs -text
4
+ *.bin filter=lfs diff=lfs merge=lfs -text
5
+ *.bz2 filter=lfs diff=lfs merge=lfs -text
6
+ *.ckpt filter=lfs diff=lfs merge=lfs -text
7
+ *.ftz filter=lfs diff=lfs merge=lfs -text
8
+ *.gz filter=lfs diff=lfs merge=lfs -text
9
+ *.h5 filter=lfs diff=lfs merge=lfs -text
10
+ *.joblib filter=lfs diff=lfs merge=lfs -text
11
+ *.lfs.* filter=lfs diff=lfs merge=lfs -text
12
+ *.lz4 filter=lfs diff=lfs merge=lfs -text
13
+ *.mds filter=lfs diff=lfs merge=lfs -text
14
+ *.mlmodel filter=lfs diff=lfs merge=lfs -text
15
+ *.model filter=lfs diff=lfs merge=lfs -text
16
+ *.msgpack filter=lfs diff=lfs merge=lfs -text
17
+ *.npy filter=lfs diff=lfs merge=lfs -text
18
+ *.npz filter=lfs diff=lfs merge=lfs -text
19
+ *.onnx filter=lfs diff=lfs merge=lfs -text
20
+ *.ot filter=lfs diff=lfs merge=lfs -text
21
+ *.parquet filter=lfs diff=lfs merge=lfs -text
22
+ *.pb filter=lfs diff=lfs merge=lfs -text
23
+ *.pickle filter=lfs diff=lfs merge=lfs -text
24
+ *.pkl filter=lfs diff=lfs merge=lfs -text
25
+ *.pt filter=lfs diff=lfs merge=lfs -text
26
+ *.pth filter=lfs diff=lfs merge=lfs -text
27
+ *.rar filter=lfs diff=lfs merge=lfs -text
28
+ *.safetensors filter=lfs diff=lfs merge=lfs -text
29
+ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
30
+ *.tar.* filter=lfs diff=lfs merge=lfs -text
31
+ *.tar filter=lfs diff=lfs merge=lfs -text
32
+ *.tflite filter=lfs diff=lfs merge=lfs -text
33
+ *.tgz filter=lfs diff=lfs merge=lfs -text
34
+ *.wasm filter=lfs diff=lfs merge=lfs -text
35
+ *.xz filter=lfs diff=lfs merge=lfs -text
36
+ *.zip filter=lfs diff=lfs merge=lfs -text
37
+ *.zst filter=lfs diff=lfs merge=lfs -text
38
+ *tfevents* filter=lfs diff=lfs merge=lfs -text
39
+ # Audio files - uncompressed
40
+ *.pcm filter=lfs diff=lfs merge=lfs -text
41
+ *.sam filter=lfs diff=lfs merge=lfs -text
42
+ *.raw filter=lfs diff=lfs merge=lfs -text
43
+ # Audio files - compressed
44
+ *.aac filter=lfs diff=lfs merge=lfs -text
45
+ *.flac filter=lfs diff=lfs merge=lfs -text
46
+ *.mp3 filter=lfs diff=lfs merge=lfs -text
47
+ *.ogg filter=lfs diff=lfs merge=lfs -text
48
+ *.wav filter=lfs diff=lfs merge=lfs -text
49
+ # Image files - uncompressed
50
+ *.bmp filter=lfs diff=lfs merge=lfs -text
51
+ *.gif filter=lfs diff=lfs merge=lfs -text
52
+ *.png filter=lfs diff=lfs merge=lfs -text
53
+ *.tiff filter=lfs diff=lfs merge=lfs -text
54
+ # Image files - compressed
55
+ *.jpg filter=lfs diff=lfs merge=lfs -text
56
+ *.jpeg filter=lfs diff=lfs merge=lfs -text
57
+ *.webp filter=lfs diff=lfs merge=lfs -text
58
+ # Video files - compressed
59
+ *.mp4 filter=lfs diff=lfs merge=lfs -text
60
+ *.webm filter=lfs diff=lfs merge=lfs -text
README.md ADDED
@@ -0,0 +1,239 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ language:
3
+ - en
4
+ task_categories:
5
+ - text-classification
6
+ annotations_creators:
7
+ - machine-generated
8
+ language_creators:
9
+ - machine-generated
10
+ source_datasets:
11
+ - original
12
+ size_categories:
13
+ - 1K<n<10K
14
+ tags:
15
+ - deception-detection
16
+ - ai-safety
17
+ - synthetic
18
+ - conversations
19
+ - llm-behavior
20
+ configs:
21
+ - config_name: default
22
+ data_files:
23
+ - split: full
24
+ path: data/full/*.parquet
25
+ ---
26
+
27
+ # SynthDeception
28
+
29
+ SynthDeception is a dataset of realistic, multi-domain, agentic scenarios for studying falsehood detection probes. Being synthetically generated from the ground up, each row provides rich metadata on how a given scenario was constructed. The main columns are "conversation_structured", which contains the full conversation of the scenario, and "follow_ups", which contains a set of possible follow-up question and answer pairs to append to the conversation. Details on all the other columns are below.
30
+
31
+ ```python
32
+ from datasets import load_dataset
33
+
34
+ ds = load_dataset("AlignmentResearch/synth-deception", split="full")
35
+ ```
36
+
37
+ The dataset has one split, `full`. Each row is one conversation, identified by `scenario_id`, and holds its 12
38
+ question-answer pairs. Rows are sorted by `scenario_id`.
39
+
40
+ ## Dataset at a glance
41
+
42
+ <p align="center">
43
+ <a href="https://huggingface.co/datasets/AlignmentResearch/synth-deception/resolve/main/assets/dataset-at-a-glance.png">
44
+ <img src="https://huggingface.co/datasets/AlignmentResearch/synth-deception/resolve/main/assets/dataset-at-a-glance.png" width="1600" style="max-width: 100%; height: auto;" alt="SynthDeception at a glance: 1,264 scenarios and 15,168 follow-up question-answer pairs (6 questions with 2 answers each per scenario), half honest and half dishonest, mirrored per question. The scenarios vary in deception type (6 ways to hide the truth or show something false), non-deception type (6 types, 4 easily mistaken for deception), information structure, deception profile, industry domain (17 fields), tool use (595 tools, 163 real and 432 simulated), user persona, and context length. The follow-ups vary in question focus (intent, belief, or common knowledge), question type (closed-ended or open-ended), question presupposition (none, established, or loaded), and answer design.">
45
+ </a>
46
+ <em>Figure 1. Overview and major axes of variation.</em>
47
+ </p>
48
+
49
+ <p align="center">
50
+ <a href="https://huggingface.co/datasets/AlignmentResearch/synth-deception/resolve/main/assets/scenario-composition-and-length.png">
51
+ <img src="https://huggingface.co/datasets/AlignmentResearch/synth-deception/resolve/main/assets/scenario-composition-and-length.png" width="1600" style="max-width: 100%; height: auto;" alt="Left: the 1,264 scenarios by type. 635 deception scenarios: hide 337 (disguising 121, camouflaging 113, confounding 103) and show 298 (distracting 109, fabricating 95, mimicking 94). 629 non-deception scenarios: confusing 419 (progressive disclosure 110, authorized dissimulation 106, misinformation 105, authorized simulation 98) and non-confusing 210 (other misbehavior 109, normal 101). Middle: turns per scenario, from 1 to 48, median 4. Right: tokens per scenario, median 6,266, up to 109k.">
52
+ </a>
53
+ <em>Figure 2. Dataset composition and statistics.</em>
54
+ </p>
55
+
56
+ ## Fields
57
+
58
+ ### Seed parameters
59
+
60
+ These fields are the scenario properties sampled before generation. They are discrete labels, and double as covariates
61
+ for per-attribute analysis. The deception fields are filled only when `deception_by_assistant` is true, and
62
+ `nondeception_type` only when it is false.
63
+
64
+ | Field | Type | Description |
65
+ |---|---|---|
66
+ | `deception_by_assistant` | `bool` | Whether the assistant is the deceiving party |
67
+ | `incomplete_information` | `bool` | Whether the scenario involves incomplete information |
68
+ | `imperfect_information` | `bool` | Whether the scenario involves imperfect/noisy information |
69
+ | `unawareness` | `bool` | Whether unknown-unknowns exist — some actors may not know what they don't know |
70
+ | `awareness_advantage` | `string` | The assistant's epistemic advantage over the user (`high`, `moderate`, `low`, `none`, `inversely_high`, `inversely_moderate`, `inversely_low`); `"NA"` when `unawareness=false` |
71
+ | `deception_effect_on_assistant` | `string` | Net effect of the deception on the assistant (`beneficial`, `neutral`, `harmful`); `"NA"` when `deception_by_assistant=false` |
72
+ | `deception_effect_on_user` | `string` | Net effect of the deception on the user (`beneficial`, `neutral`, `harmful`); `"NA"` when `deception_by_assistant=false` |
73
+ | `deception_obviousness` | `string` | How apparent the deception is within the conversation (`low`, `moderate`, `high`); `"NA"` when `deception_by_assistant=false` |
74
+ | `deception_verifiability` | `string` | How easily the deception can be externally verified (`low`, `moderate`, `high`); `"NA"` when `deception_by_assistant=false` |
75
+ | `deception_suspected` | `bool` (nullable) | Whether the user suspects deception during the conversation; null when `deception_by_assistant=false` |
76
+ | `deception_type` | `string` | Structural category of the deception (see below); `"NA"` when `deception_by_assistant=false` |
77
+ | `nondeception_type` | `string` | Category of non-deception behavior (see below); `"NA"` when `deception_by_assistant=true` |
78
+ | `data_domain` | `string` | Thematic domain of the conversation (see below) |
79
+ | `tool_calls` | `bool` | Whether the assistant uses tool calls in the conversation |
80
+ | `conversation_length` | `string` | Approximate length of the conversation (`very_short`, `short`, `moderate`, `long`, `very_long`) |
81
+ | `system_prompt_length` | `string` | Approximate length category of the assembled system prompt |
82
+
83
+ ### Deception types
84
+
85
+ When `deception_by_assistant=true`, `deception_type` is drawn from the Bell & Whaley taxonomy:
86
+
87
+ | Value | Category | Description |
88
+ |---|---|---|
89
+ | `show_fabricating` | Simulation | The assistant presents invented facts or data as if they were real |
90
+ | `show_distracting` | Simulation | The assistant draws attention to irrelevant information to divert from the truth |
91
+ | `show_mimicking` | Simulation | The assistant presents false content that closely mimics legitimate patterns |
92
+ | `hide_camouflaging` | Dissimulation | The assistant conceals true information by blending it with surrounding content |
93
+ | `hide_confounding` | Dissimulation | The assistant obscures the truth by adding misleading context or conflating concepts |
94
+ | `hide_disguising` | Dissimulation | The assistant disguises true information so it appears to be something else |
95
+
96
+ ### Non-deception types
97
+
98
+ When `deception_by_assistant=false`, `nondeception_type` categorizes the honest scenario. Four of the six types are
99
+ *confusing non-deception* — behaviors that may superficially resemble deception but are categorically distinct:
100
+
101
+ | Value | Category | Description |
102
+ |---|---|---|
103
+ | `normal` | Conventional | Standard honest behavior — providing information believed to be true and relevant |
104
+ | `other_misbehavior` | Conventional | Misbehavior that is not deception (e.g., unhelpfulness, errors) |
105
+ | `misinformation` | Confusing | Communicating things believed to be true that are actually false — no intent to deceive |
106
+ | `authorized_dissimulation` | Confusing | Withholding information the other party is not expected to receive given context (e.g., confidentiality) |
107
+ | `authorized_simulation` | Confusing | Communicating things not meant to be taken literally (e.g., roleplay, sarcasm) |
108
+ | `progressive_disclosure` | Confusing | Honest partial disclosure with the opportunity to request more detail later |
109
+
110
+ ### Data domains
111
+
112
+ Conversations span 17 domains: `biological`, `chemical`, `radiological`, `nuclear`, `energy`, `water`, `agriculture`,
113
+ `transportation`, `cybersecurity`, `communications`, `finance`, `economics`, `healthcare`, `education`, `politics`,
114
+ `ai_safety`, `personal_assistance`.
115
+
116
+ ### Generated scenario and context fields
117
+
118
+ | Field | Type | Description |
119
+ |---|---|---|
120
+ | `user_persona_base` | `struct` | Sampled base persona for the user (demographics, personality traits) |
121
+ | `scenario_outline` | `string` | Narrative description of the scenario |
122
+ | `scenario_outline_structured` | `struct` | Structured scenario with per-turn outlines (`overview`, `conversation_turns`, `available_tools`) |
123
+ | `assistant_persona_instructions` | `string` | System-level persona instructions for the assistant |
124
+ | `assistant_task_instructions` | `string` | Task-level instructions defining the assistant's role and duties |
125
+ | `user_persona_instructions` | `string` | Instructions for the user's persona used during conversation generation |
126
+ | `user_description` | `string` | User context passage injected into the system prompt |
127
+ | `historical_context` | `string` | Simulated prior-interaction history provided to the assistant |
128
+ | `domain_context` | `string` | Domain background knowledge injected into the assistant context |
129
+ | `system_prompt` | `string` | Full assembled system prompt used when generating the conversation |
130
+
131
+ ### Conversation
132
+
133
+ | Field | Type | Description |
134
+ |---|---|---|
135
+ | `conversation_structured` | `struct` | The generated conversation, with `messages` and `tools`. Each message has a `role` (`system`, `user`, `assistant`, or `tool`) and `content`; assistant messages that call tools carry `tool_calls`, and `tool`-role messages carry a `tool_call_id`. `tools` lists the tool/function definitions available in the conversation (empty when the assistant uses no tools). |
136
+
137
+ ### Follow-up interrogation
138
+
139
+ | Field | Type | Description |
140
+ |---|---|---|
141
+ | `follow_ups` | `struct` | `follow_up_groups`: **6 groups** = 3 interrogation foci (`intent`, `belief`, `common knowledge`) × 2 question types (`follow_up_type` ∈ `closed-ended`, `open-ended`). Each group has a list of `follow_ups` items, each item being `{question, honest_answer, dishonest_answer}` plus the taxonomy fields below — the contrasting honest/dishonest answers are the ground truth for probing introspective honesty. |
142
+
143
+ Each item also carries the question's taxonomy:
144
+
145
+ | Field | Description |
146
+ |---|---|
147
+ | `presupposition` | `none`: the question takes nothing for granted that the answer could confirm or correct. `established`: the question takes for granted only what the conversation itself put on the table, which may be false if the assistant asserted it. `loaded`: the question takes for granted the very point it is probing, so that answering it directly concedes that point |
148
+ | `honest_stance`, `dishonest_stance` | What each answer does with that presupposition. `accepts`: the answer stays inside the presupposition and answers the question. `rejects`: the answer denies the presupposition instead of answering within it, using the mandated fixed shape: the minimal denial, then a comma, then exactly one clause stating what happened instead (e.g. "Neither, I escalated it to the review queue."). `NA` exactly when there is none |
149
+ | `closed_ended_form` | Closed-ended questions only. Polar is yes/no; alternative offers a choice |
150
+ | `option_count` | How many options the question offers |
151
+ | `honest_option_position` | The drawn float *q*, in [0, 1]. The 1-based option index the honest answer must take is `min(floor(q * n), n - 1) + 1` |
152
+
153
+ The columns `follow_up_<field>_<focus>_<closed|open>_q1` repeat `presupposition`, `honest_stance`, `dishonest_stance`,
154
+ `closed_ended_form`, and `honest_option_position` as flat columns, one per group. `follow_up_rewrite_direction` is the
155
+ concrete rewrite direction for the row: `honest_to_dishonest` or `dishonest_to_honest`.
156
+
157
+ ## How the data was made
158
+
159
+ All data is synthetic. The scenarios, the conversations, and the follow-up questions were written by
160
+ language models, and a repair pass later corrected part of them (see the last paragraph). The user
161
+ personas come from NVIDIA's Nemotron-Personas dataset of synthetic person records.
162
+
163
+ <p align="center">
164
+ <a href="https://huggingface.co/datasets/AlignmentResearch/synth-deception/resolve/main/assets/synthdeception-dag.png">
165
+ <img src="https://huggingface.co/datasets/AlignmentResearch/synth-deception/resolve/main/assets/synthdeception-dag.png" width="1600" style="max-width: 100%; height: auto;" alt="How the SynthDeception columns are made: sampled seed parameters feed the scenario outline, which feeds the context texts, the system prompt, the conversation, and the follow-ups.">
166
+ </a>
167
+ <em>Figure 3. How each column is generated from the others.</em>
168
+ </p>
169
+
170
+ Each row was generated by a synthetic data generation (SDG) pipeline in five steps:
171
+
172
+ 1. **Seed parameters.** Each seed column is drawn from a fixed list of values. The six deception columns
173
+ are drawn only when `deception_by_assistant` is true, `nondeception_type` only when it is false, and
174
+ `awareness_advantage` only when `unawareness` is true. `user_persona_base` is a person record drawn from
175
+ Nemotron-Personas (US locale, ages 15 to 75). The follow-up control columns (`follow_up_*_q1` and
176
+ `follow_up_rewrite_direction`) are drawn here too.
177
+ 2. **Scenario outline.** A frontier model writes the outline from the seed parameters. Before it writes,
178
+ it can look up stable facts with up to five read-only real tools, and a summary of those facts is added
179
+ to its prompt. The outline is an overview, a turn-by-turn plan of what the user and the assistant say
180
+ in each turn with an estimated word count, and the tools available to the assistant (real or
181
+ simulated). `scenario_outline_structured` is the parsed form of the same output.
182
+ 3. **Context texts.** The frontier model writes six texts, each in a separate call that reads the
183
+ outline: `assistant_persona_instructions`, `assistant_task_instructions`, `domain_context`,
184
+ `historical_context`, `user_persona_instructions` (which also reads `user_persona_base`), and
185
+ `user_description` (which also reads `user_persona_instructions`). Some calls also read `tool_calls`
186
+ and `system_prompt_length`. `system_prompt` is not generated: a fixed template joins five of these
187
+ texts (all except `user_persona_instructions`).
188
+ 4. **Conversation.** The conversation renders the outline's turn-by-turn plan. For each turn, one model
189
+ call writes the user's message and a second writes the assistant's. Each call gets the current turn's
190
+ description from the plan, the descriptions and the text of the previous turns, and its own context:
191
+ `user_persona_instructions` for the user, `system_prompt` for the assistant. Both are instructed to
192
+ render the description faithfully, without softening, omitting, adding to, or correcting the planned
193
+ behavior. When `tool_calls` is true and the outline lists tools, the assistant's turn can call them.
194
+ Real tools are executed. Simulated tools return a result written by a model.
195
+ 5. **Follow-ups.** The frontier model writes the follow-up questions from the outline, the follow-up
196
+ control columns, and `deception_by_assistant`. It does not read the conversation. A first call drafts
197
+ 6 groups (3 interrogation foci × 2 question types), and each question gets an honest and a dishonest
198
+ answer that match the drawn control values. A second call rewrites both answers of each question from
199
+ the side named in `follow_up_rewrite_direction`, so that the two answers share their surface features.
200
+
201
+ A model judge checked each generated text as it was produced: the outline, each context text, the
202
+ conversation, each simulated tool result, and the follow-ups. A rejected text was regenerated with the
203
+ judge's reason added to the prompt, and the row was dropped when the retry budget ran out.
204
+
205
+ After generation, every row was audited against its seed parameters and its outline. Text that
206
+ contradicted them was corrected in the outlines, context texts, conversations, tool results, and
207
+ follow-ups. Some rows that could not be made consistent were removed. On part of the rows,
208
+ the follow-up answers were also rewritten to reduce word-level cues that separate honest from dishonest
209
+ answers.
210
+
211
+ ## Citation
212
+
213
+ ```bibtex
214
+ @dataset{farai_synth_deception,
215
+ author = {FAR.AI Deception Pod},
216
+ title = {SynthDeception},
217
+ year = {2026},
218
+ publisher = {Hugging Face},
219
+ url = {https://huggingface.co/datasets/AlignmentResearch/synth-deception}
220
+ }
221
+
222
+ @misc{farai_synth_deception_probes,
223
+ author = {FAR.AI Deception Pod},
224
+ title = {SynthDeception probes},
225
+ year = {2026},
226
+ publisher = {Hugging Face},
227
+ url = {https://huggingface.co/AlignmentResearch/synth-deception-probes}
228
+ }
229
+
230
+ @misc{hollinsworth2026tellingfibsprobeseffectively,
231
+ title={Telling FIBS: Probes Effectively Detect Sabotage and Catch Unverbalized Deception},
232
+ author={Oskar J. Hollinsworth and Alex F. Spies and Tigist Diriba and Adam Gleave and Chris Cundy},
233
+ year={2026},
234
+ eprint={2610.12445},
235
+ archivePrefix={arXiv},
236
+ primaryClass={cs.LG},
237
+ url={https://arxiv.org/abs/2610.12445},
238
+ }
239
+ ```
assets/dataset-at-a-glance.png ADDED

Git LFS Details

  • SHA256: 664c5d3cbaf8a6ae0e424671875d97be4a745fe9bffeb0014340ef62cdd03995
  • Pointer size: 131 Bytes
  • Size of remote file: 414 kB
assets/scenario-composition-and-length.png ADDED

Git LFS Details

  • SHA256: 157becaeb9027af74ce3286057b6dff90c7b1c27af4b95712e8fa771337711a2
  • Pointer size: 131 Bytes
  • Size of remote file: 513 kB
assets/synthdeception-dag.png ADDED

Git LFS Details

  • SHA256: 1a74ecc763abe6efbe2f0d31e0c03b22df2e438722f824dd675966b83b64f06d
  • Pointer size: 131 Bytes
  • Size of remote file: 207 kB
data/full/00001.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:e2f6fa5312c63a8dd9e025fea4bbe15741cf84c45348f77e44fe52762d33f048
3
+ size 72652390