Adam-12-0 commited on
Commit
26e5602
·
verified ·
1 Parent(s): ad498a1

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +61 -1
README.md CHANGED
@@ -11,6 +11,11 @@ task_ids:
11
  - multi-class-classification
12
  size_categories:
13
  - n<1K
 
 
 
 
 
14
  tags:
15
  - theory-of-mind
16
  - benchmark
@@ -21,58 +26,78 @@ tags:
21
  - mental-state-representation
22
  ---
23
 
 
24
  # Dataset Card for OmniToM
25
 
 
26
  ## Dataset Details
27
 
 
28
  ### Dataset Description
29
 
 
30
  OmniToM is a benchmark for evaluating Theory of Mind in language models through explicit belief-structure modeling. Instead of scoring only endpoint answers to social-reasoning questions, OmniToM exposes the intermediate belief structure that a model must build in order to reason coherently about what different actors know, believe, infer, intend, or misunderstand.
31
 
 
32
  Each example is a short English story paired with:
33
 
 
34
  - a set of actor-centered belief propositions
35
  - a reserved `world` actor for narrator/world facts
36
  - a seven-dimensional schema label vector for every belief
37
 
 
38
  The benchmark supports two linked tasks:
39
 
 
40
  1. **Belief Extraction**
41
  - Given a story, extract the relevant belief structure as `(Actor, Belief, Order)` tuples.
42
  2. **Belief Labeling**
43
  - Given the story and belief tuples, label each belief along seven closed-set schema dimensions.
44
 
 
45
  - **Language(s):** English
46
  - **License:** MIT
47
 
 
48
  ## Uses
49
 
 
50
  ### Direct Use
51
 
 
52
  OmniToM is designed for benchmark evaluation and diagnostic analysis. Suitable uses include:
53
 
 
54
  - zero-shot belief extraction
55
  - zero-shot belief labeling
56
  - semantic-judge evaluation of extracted belief tables
57
  - analysis of multi-actor and higher-order Theory-of-Mind reasoning
58
  - process-sensitive evaluation beyond endpoint question answering
59
 
 
60
  ### Out-of-Scope Use
61
 
 
62
  OmniToM should not be treated as:
63
 
 
64
  - a direct measure of real-world social intelligence
65
  - a measure of embodied, interactive, or multimodal social reasoning
66
  - a safety certification benchmark for deployed systems
67
  - a clinical, educational, or psychological assessment tool
68
  - a complete coverage benchmark for all possible Theory-of-Mind phenomena
69
 
 
70
  ## Dataset Structure
71
 
 
72
  ### Data Instances
73
 
 
74
  Each line in the release file is one JSON object:
75
 
 
76
  ```json
77
  {
78
  "story_id": 1,
@@ -96,8 +121,10 @@ Each line in the release file is one JSON object:
96
  }
97
  ```
98
 
 
99
  ### Data Fields
100
 
 
101
  - `story_id`
102
  - Unique integer identifier for the story.
103
  - `story_category`
@@ -125,26 +152,37 @@ Each line in the release file is one JSON object:
125
  - `beliefs[].labels.context`
126
  - `Neutral`, `Temporal`, `Deceptive`, or `Counterfactual`.
127
 
 
128
  ### Data Splits
129
 
 
130
  This release contains one benchmark split:
131
 
 
132
  - `train` / benchmark split: `895` stories
133
 
 
134
  The split is named `train` in the Hugging Face dataset viewer for compatibility with the default dataset loading interface. It should be interpreted as the benchmark split, not as a recommended training set.
135
 
 
136
  ## Dataset Creation
137
 
 
138
  ### Curation Rationale
139
 
 
140
  OmniToM was created to address a limitation in prior Theory-of-Mind benchmarks for language models: most evaluate endpoint question answering rather than whether a model constructs a coherent belief representation while reading the story.
141
 
 
142
  OmniToM instead evaluates explicit belief-structure modeling. The benchmark is grounded in short stories from ToMBench and organizes reasoning around an ATOMS-grounded belief-level schema for fine-grained analysis of mental-state representations.
143
 
 
144
  ### Source Data
145
 
 
146
  The benchmark sources its stories from ToMBench. From the original source corpus, OmniToM retains seven story categories whose stories provide sufficiently self-contained mental-state evidence for belief extraction from text alone:
147
 
 
148
  - Ambiguous Story Task
149
  - False Belief Task
150
  - Faux-pas Recognition Test
@@ -153,37 +191,50 @@ The benchmark sources its stories from ToMBench. From the original source corpus
153
  - Scalar Implicature Test
154
  - Strange Story Task
155
 
 
156
  ### Annotation Process
157
 
 
158
  The benchmark was built with a human-validated, LLM-assisted annotation pipeline:
159
 
 
160
  - `1,383` source stories in the original corpus
161
  - `916` stories retained after source filtering
162
  - `895` stories released in the final benchmark
163
  - `22,343` labeled belief propositions in the released benchmark
164
  - `156,401` total schema labels in the released benchmark
165
 
 
166
  The accompanying paper reports:
167
 
 
168
  - Stage 1 expert-overlap validation after reconciliation: `83.72%`
169
  - Stage 2 strict all-annotator exact-match label reliability: `92.38%`
170
  - Human-human agreement on the semantic-alignment validation set: `88.86%`
171
  - Human-judge agreement for the selected semantic judge: `87.76%`
172
 
 
173
  ### Who are the source data producers?
174
 
 
175
  The story texts are sourced from ToMBench. The belief structures and schema labels are benchmark annotations produced through the OmniToM human-validated annotation pipeline described in the accompanying paper.
176
 
 
177
  ### Personal and Sensitive Information
178
 
 
179
  The release consists of short benchmark stories and belief annotations. It is not designed to contain personal user data, private communications, direct identifiers, medical records, or real-world sensitive records. As with many story-based datasets, names, family roles, occupations, emotions, intentions, and social situations may appear in the source material, but the benchmark is intended for research evaluation rather than identification or profiling.
180
 
 
181
  ## Bias, Risks, and Limitations
182
 
 
183
  OmniToM is story-based, text-only, English-language, and sourced from a specific upstream benchmark distribution. It therefore reflects the representational biases, writing conventions, scenario distribution, and coverage limitations of its source stories. Names, roles, social situations, and pragmatic conventions may also be unevenly distributed across categories.
184
 
 
185
  Additional known limitations:
186
 
 
187
  - the benchmark evaluates story-based Theory of Mind rather than interactive, embodied, or multimodal social reasoning
188
  - the retained stories are short and self-contained, and do not stress-test long-horizon information tracking, dense temporal structure, or deeply nested mental states beyond the order-3 schema
189
  - the released labels come from a human-validated LLM-assisted pipeline rather than fully manual annotation of every story
@@ -191,12 +242,16 @@ Additional known limitations:
191
  - Stage 1 extraction evaluation in the paper relies on a human-validated semantic judge rather than full human adjudication across the full release
192
  - the selected semantic judge reached `87.76%` agreement with human semantic-alignment decisions, so extraction `F1` should be interpreted as an approximate aggregate metric rather than an exact belief-level alignment score
193
 
 
194
  ### Recommendations
195
 
 
196
  Users should interpret OmniToM as a diagnostic benchmark for explicit belief-structure modeling. Benchmark scores should not be treated as evidence of robust real-world interpersonal reasoning, embodied social competence, clinical or educational validity, or deployment safety.
197
 
 
198
  ## Citation
199
 
 
200
  ```bibtex
201
  @misc{omnitom2026,
202
  title={OmniToM: Benchmarking Theory of Mind in LLMs via Explicit Belief Modeling},
@@ -207,10 +262,15 @@ Users should interpret OmniToM as a diagnostic benchmark for explicit belief-str
207
  }
208
  ```
209
 
 
210
  ## More Information
211
 
 
212
  Paper: [OmniToM on arXiv](https://arxiv.org/abs/2605.26322)
213
 
214
- Code and evaluation scripts: [Adam-12-0/omnitom-benchmark](https://github.com/Adam-12-0/omnitom-benchmark)
 
 
215
 
216
  Project page: [OmniToM](https://adam-12-0.github.io/omnitom-project/)
 
 
11
  - multi-class-classification
12
  size_categories:
13
  - n<1K
14
+ configs:
15
+ - config_name: default
16
+ data_files:
17
+ - split: train
18
+ path: benchmark_story_belief_labels.jsonl
19
  tags:
20
  - theory-of-mind
21
  - benchmark
 
26
  - mental-state-representation
27
  ---
28
 
29
+
30
  # Dataset Card for OmniToM
31
 
32
+
33
  ## Dataset Details
34
 
35
+
36
  ### Dataset Description
37
 
38
+
39
  OmniToM is a benchmark for evaluating Theory of Mind in language models through explicit belief-structure modeling. Instead of scoring only endpoint answers to social-reasoning questions, OmniToM exposes the intermediate belief structure that a model must build in order to reason coherently about what different actors know, believe, infer, intend, or misunderstand.
40
 
41
+
42
  Each example is a short English story paired with:
43
 
44
+
45
  - a set of actor-centered belief propositions
46
  - a reserved `world` actor for narrator/world facts
47
  - a seven-dimensional schema label vector for every belief
48
 
49
+
50
  The benchmark supports two linked tasks:
51
 
52
+
53
  1. **Belief Extraction**
54
  - Given a story, extract the relevant belief structure as `(Actor, Belief, Order)` tuples.
55
  2. **Belief Labeling**
56
  - Given the story and belief tuples, label each belief along seven closed-set schema dimensions.
57
 
58
+
59
  - **Language(s):** English
60
  - **License:** MIT
61
 
62
+
63
  ## Uses
64
 
65
+
66
  ### Direct Use
67
 
68
+
69
  OmniToM is designed for benchmark evaluation and diagnostic analysis. Suitable uses include:
70
 
71
+
72
  - zero-shot belief extraction
73
  - zero-shot belief labeling
74
  - semantic-judge evaluation of extracted belief tables
75
  - analysis of multi-actor and higher-order Theory-of-Mind reasoning
76
  - process-sensitive evaluation beyond endpoint question answering
77
 
78
+
79
  ### Out-of-Scope Use
80
 
81
+
82
  OmniToM should not be treated as:
83
 
84
+
85
  - a direct measure of real-world social intelligence
86
  - a measure of embodied, interactive, or multimodal social reasoning
87
  - a safety certification benchmark for deployed systems
88
  - a clinical, educational, or psychological assessment tool
89
  - a complete coverage benchmark for all possible Theory-of-Mind phenomena
90
 
91
+
92
  ## Dataset Structure
93
 
94
+
95
  ### Data Instances
96
 
97
+
98
  Each line in the release file is one JSON object:
99
 
100
+
101
  ```json
102
  {
103
  "story_id": 1,
 
121
  }
122
  ```
123
 
124
+
125
  ### Data Fields
126
 
127
+
128
  - `story_id`
129
  - Unique integer identifier for the story.
130
  - `story_category`
 
152
  - `beliefs[].labels.context`
153
  - `Neutral`, `Temporal`, `Deceptive`, or `Counterfactual`.
154
 
155
+
156
  ### Data Splits
157
 
158
+
159
  This release contains one benchmark split:
160
 
161
+
162
  - `train` / benchmark split: `895` stories
163
 
164
+
165
  The split is named `train` in the Hugging Face dataset viewer for compatibility with the default dataset loading interface. It should be interpreted as the benchmark split, not as a recommended training set.
166
 
167
+
168
  ## Dataset Creation
169
 
170
+
171
  ### Curation Rationale
172
 
173
+
174
  OmniToM was created to address a limitation in prior Theory-of-Mind benchmarks for language models: most evaluate endpoint question answering rather than whether a model constructs a coherent belief representation while reading the story.
175
 
176
+
177
  OmniToM instead evaluates explicit belief-structure modeling. The benchmark is grounded in short stories from ToMBench and organizes reasoning around an ATOMS-grounded belief-level schema for fine-grained analysis of mental-state representations.
178
 
179
+
180
  ### Source Data
181
 
182
+
183
  The benchmark sources its stories from ToMBench. From the original source corpus, OmniToM retains seven story categories whose stories provide sufficiently self-contained mental-state evidence for belief extraction from text alone:
184
 
185
+
186
  - Ambiguous Story Task
187
  - False Belief Task
188
  - Faux-pas Recognition Test
 
191
  - Scalar Implicature Test
192
  - Strange Story Task
193
 
194
+
195
  ### Annotation Process
196
 
197
+
198
  The benchmark was built with a human-validated, LLM-assisted annotation pipeline:
199
 
200
+
201
  - `1,383` source stories in the original corpus
202
  - `916` stories retained after source filtering
203
  - `895` stories released in the final benchmark
204
  - `22,343` labeled belief propositions in the released benchmark
205
  - `156,401` total schema labels in the released benchmark
206
 
207
+
208
  The accompanying paper reports:
209
 
210
+
211
  - Stage 1 expert-overlap validation after reconciliation: `83.72%`
212
  - Stage 2 strict all-annotator exact-match label reliability: `92.38%`
213
  - Human-human agreement on the semantic-alignment validation set: `88.86%`
214
  - Human-judge agreement for the selected semantic judge: `87.76%`
215
 
216
+
217
  ### Who are the source data producers?
218
 
219
+
220
  The story texts are sourced from ToMBench. The belief structures and schema labels are benchmark annotations produced through the OmniToM human-validated annotation pipeline described in the accompanying paper.
221
 
222
+
223
  ### Personal and Sensitive Information
224
 
225
+
226
  The release consists of short benchmark stories and belief annotations. It is not designed to contain personal user data, private communications, direct identifiers, medical records, or real-world sensitive records. As with many story-based datasets, names, family roles, occupations, emotions, intentions, and social situations may appear in the source material, but the benchmark is intended for research evaluation rather than identification or profiling.
227
 
228
+
229
  ## Bias, Risks, and Limitations
230
 
231
+
232
  OmniToM is story-based, text-only, English-language, and sourced from a specific upstream benchmark distribution. It therefore reflects the representational biases, writing conventions, scenario distribution, and coverage limitations of its source stories. Names, roles, social situations, and pragmatic conventions may also be unevenly distributed across categories.
233
 
234
+
235
  Additional known limitations:
236
 
237
+
238
  - the benchmark evaluates story-based Theory of Mind rather than interactive, embodied, or multimodal social reasoning
239
  - the retained stories are short and self-contained, and do not stress-test long-horizon information tracking, dense temporal structure, or deeply nested mental states beyond the order-3 schema
240
  - the released labels come from a human-validated LLM-assisted pipeline rather than fully manual annotation of every story
 
242
  - Stage 1 extraction evaluation in the paper relies on a human-validated semantic judge rather than full human adjudication across the full release
243
  - the selected semantic judge reached `87.76%` agreement with human semantic-alignment decisions, so extraction `F1` should be interpreted as an approximate aggregate metric rather than an exact belief-level alignment score
244
 
245
+
246
  ### Recommendations
247
 
248
+
249
  Users should interpret OmniToM as a diagnostic benchmark for explicit belief-structure modeling. Benchmark scores should not be treated as evidence of robust real-world interpersonal reasoning, embodied social competence, clinical or educational validity, or deployment safety.
250
 
251
+
252
  ## Citation
253
 
254
+
255
  ```bibtex
256
  @misc{omnitom2026,
257
  title={OmniToM: Benchmarking Theory of Mind in LLMs via Explicit Belief Modeling},
 
262
  }
263
  ```
264
 
265
+
266
  ## More Information
267
 
268
+
269
  Paper: [OmniToM on arXiv](https://arxiv.org/abs/2605.26322)
270
 
271
+
272
+ Code and evaluation scripts: [BridgeAI-Lab/OMNITOM](https://github.com/BridgeAI-Lab/OMNITOM)
273
+
274
 
275
  Project page: [OmniToM](https://adam-12-0.github.io/omnitom-project/)
276
+