7jep7 commited on
Commit
b5b05db
·
verified ·
1 Parent(s): 5da4fa8

Deanonymised release: card, knowledge graph, Lite subset

Browse files
.gitattributes CHANGED
@@ -69,3 +69,12 @@ factorybench_qa/level_3/test.jsonl filter=lfs diff=lfs merge=lfs -text
69
  factorybench_qa/level_4/train.jsonl filter=lfs diff=lfs merge=lfs -text
70
  factorybench_qa/level_4/validation.jsonl filter=lfs diff=lfs merge=lfs -text
71
  factorybench_qa/level_4/test.jsonl filter=lfs diff=lfs merge=lfs -text
 
 
 
 
 
 
 
 
 
 
69
  factorybench_qa/level_4/train.jsonl filter=lfs diff=lfs merge=lfs -text
70
  factorybench_qa/level_4/validation.jsonl filter=lfs diff=lfs merge=lfs -text
71
  factorybench_qa/level_4/test.jsonl filter=lfs diff=lfs merge=lfs -text
72
+ factorybench_lite/level_2/train.jsonl filter=lfs diff=lfs merge=lfs -text
73
+ factorybench_lite/level_3/train.jsonl filter=lfs diff=lfs merge=lfs -text
74
+ factorybench_qa/level_1.jsonl filter=lfs diff=lfs merge=lfs -text
75
+ factorybench_qa/level_2.jsonl filter=lfs diff=lfs merge=lfs -text
76
+ factorybench_qa/level_3.jsonl filter=lfs diff=lfs merge=lfs -text
77
+ factorybench_qa/level_4.jsonl filter=lfs diff=lfs merge=lfs -text
78
+ factorybench_lite/level_2.jsonl filter=lfs diff=lfs merge=lfs -text
79
+ factorybench_lite/level_3.jsonl filter=lfs diff=lfs merge=lfs -text
80
+ factorybench_lite/level_4.jsonl filter=lfs diff=lfs merge=lfs -text
README.md CHANGED
@@ -4,6 +4,7 @@ language:
4
  - en
5
  task_categories:
6
  - question-answering
 
7
  tags:
8
  - robotics
9
  - industrial
@@ -14,6 +15,39 @@ tags:
14
  pretty_name: FactoryBench
15
  size_categories:
16
  - 10K<n<100K
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
17
  ---
18
 
19
  # FactoryBench
@@ -31,7 +65,8 @@ The benchmark is grounded in **FactoryWave**, a dense multivariate telemetry dat
31
 
32
  ## Dataset summary
33
 
34
- - **70,918 Q&A pairs** across four causal levels and three splits (train/validation/test).
 
35
  - **5 answer formats**: single-select MCQ, multi-select MCQ, ranking, tensor/numerical, free-form (judged by an LLM-as-judge voting protocol).
36
  - **Telemetry from real industrial robots** with systematic fault injection (27 atomic mechanisms across pick-and-place, screwing, and peg-in-hole tasks).
37
 
@@ -39,26 +74,65 @@ The benchmark is grounded in **FactoryWave**, a dense multivariate telemetry dat
39
 
40
  ```
41
  FactoryBench/
42
- └── factorybench_qa/ # Question-answer pairs
43
- ├── level_1/{train,validation,test}.jsonl
44
- ├── level_2/{train,validation,test}.jsonl
45
- ├── level_3/{train,validation,test}.jsonl
46
- └── level_4/{train,validation,test}.jsonl
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
47
  ```
48
 
49
  ## Q&A pair counts
50
 
51
- | Level | Train | Val | Test | Total |
52
- |-------|--------|-------|-------|---------|
53
- | L1 | 12,674 | 1,338 | 1,309 | 15,321 |
54
- | L2 | 33,311 | 3,428 | 3,487 | 40,226 |
55
- | L3 | 2,353 | 265 | 321 | 2,939 |
56
- | L4 | 9,949 | 1,251 | 1,232 | 12,432 |
57
- | **Total** | **58,287** | **6,282** | **6,349** | **70,918** |
 
 
 
 
 
58
 
59
  ## Q&A fields
60
 
61
- Each line in `factorybench_qa/level_*/*.jsonl` is a single Q&A item:
62
 
63
  | Field | Description |
64
  |-------|-------------|
@@ -80,12 +154,33 @@ Each line in `factorybench_qa/level_*/*.jsonl` is a single Q&A item:
80
  ```python
81
  from datasets import load_dataset
82
 
83
- # Load a single level/split
 
 
 
 
 
 
84
  ds = load_dataset(
85
  "Forgis/FactoryBench",
86
- data_files="factorybench_qa/level_1/test.jsonl",
87
  split="train",
88
  )
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
89
  ```
90
 
91
  ## Citation
@@ -93,25 +188,28 @@ ds = load_dataset(
93
  If you use FactoryBench, please cite the dataset and the two upstream open-source datasets it incorporates (AURSAD and voraus-AD):
94
 
95
  ```bibtex
96
- @misc{anonymous2026factorybench,
97
- title = {FactoryBench: Evaluating Industrial Machine Understanding},
98
- author = {Anonymous},
99
- year = {2026},
100
- note = {Submission under double-blind review}
 
 
 
101
  }
102
 
103
  @article{leporowski2022aursad,
104
  title = {{AURSAD}: Universal Robot Screwdriving Anomaly Detection Dataset},
105
- author = {Leporowski, B{\l}a{\.z}ej and Tola, Daniella and Hansen, Christian and Iosifidis, Alexandros},
106
- journal = {arXiv preprint arXiv:2202.03211},
107
- year = {2022}
108
  }
109
 
110
  @misc{brockmann2024vorausad,
111
  title = {voraus-{AD}: A New Dataset for Anomaly Detection in Robot Applications},
112
  author = {Brockmann, Jan Thie{\ss} and Rudolph, Marco and Rosenhahn, Bodo and Wandt, Bastian},
113
  year = {2024},
114
- eprint = {2311.04153},
115
  archivePrefix = {arXiv},
116
  primaryClass = {cs.RO}
117
  }
@@ -127,4 +225,4 @@ If you use FactoryBench, please cite the dataset and the two upstream open-sourc
127
 
128
  ## License
129
 
130
- Released under the [Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0)](https://creativecommons.org/licenses/by-nc/4.0/) license. You may share and adapt the dataset for non-commercial purposes with appropriate attribution.
 
4
  - en
5
  task_categories:
6
  - question-answering
7
+ - time-series-forecasting
8
  tags:
9
  - robotics
10
  - industrial
 
15
  pretty_name: FactoryBench
16
  size_categories:
17
  - 10K<n<100K
18
+ configs:
19
+ - config_name: level_1
20
+ data_files:
21
+ - split: train
22
+ path: factorybench_qa/level_1.jsonl
23
+ - config_name: level_2
24
+ data_files:
25
+ - split: train
26
+ path: factorybench_qa/level_2.jsonl
27
+ - config_name: level_3
28
+ data_files:
29
+ - split: train
30
+ path: factorybench_qa/level_3.jsonl
31
+ - config_name: level_4
32
+ data_files:
33
+ - split: train
34
+ path: factorybench_qa/level_4.jsonl
35
+ - config_name: lite_level_1
36
+ data_files:
37
+ - split: train
38
+ path: factorybench_lite/level_1.jsonl
39
+ - config_name: lite_level_2
40
+ data_files:
41
+ - split: train
42
+ path: factorybench_lite/level_2.jsonl
43
+ - config_name: lite_level_3
44
+ data_files:
45
+ - split: train
46
+ path: factorybench_lite/level_3.jsonl
47
+ - config_name: lite_level_4
48
+ data_files:
49
+ - split: train
50
+ path: factorybench_lite/level_4.jsonl
51
  ---
52
 
53
  # FactoryBench
 
65
 
66
  ## Dataset summary
67
 
68
+ - **69,691 Q&A pairs** across four causal levels, released as a single undivided pool (the public release is not split into train/validation/test).
69
+ - **FactoryBench-Lite**: a balanced 3,000-item subset for cheap evaluation, even across templates and within each template on the dimension that determines its answer.
70
  - **5 answer formats**: single-select MCQ, multi-select MCQ, ranking, tensor/numerical, free-form (judged by an LLM-as-judge voting protocol).
71
  - **Telemetry from real industrial robots** with systematic fault injection (27 atomic mechanisms across pick-and-place, screwing, and peg-in-hole tasks).
72
 
 
74
 
75
  ```
76
  FactoryBench/
77
+ ├── factorybench_qa/ # Question-answer pairs (full pool)
78
+ │ ├── level_1.jsonl
79
+ │ ├── level_2.jsonl
80
+ │ ├── level_3.jsonl
81
+ │ └── level_4.jsonl
82
+ ├── factorybench_lite/ # Balanced 3,000-item evaluation subset
83
+ │ ├── level_1.jsonl
84
+ │ ├── level_2.jsonl
85
+ │ ├── level_3.jsonl
86
+ │ └── level_4.jsonl
87
+ ├── knowledge_graph/ # Combined knowledge graph
88
+ │ ├── knowledge_graph.json # Machines, grippers, tasks, events, faults, anomalies, error→protocol map, relevance specs
89
+ │ └── SCHEMA.md # Field-level schema documentation
90
+ └── factorywave/ # Underlying telemetry & metadata
91
+ ├── episodes.parquet # Episode-level metadata (9,728 episodes)
92
+ ├── flow.parquet # Task flow definitions
93
+ ├── kuka_signals.parquet # KUKA KR10 signals (~83 Hz, 1,428 episodes)
94
+ ├── ur_signals.parquet # UR3 signals (~125 Hz, 3,076 episodes)
95
+ ├── ur_signals_10hz.parquet # UR3 signals (10 Hz, 3,984 episodes — disjoint from ur_signals)
96
+ └── ur_screwdriver_signals.parquet # UR3 screwdriver subset (~125 Hz, 1,240 episodes)
97
+ ```
98
+
99
+ > **Note on UR coverage.** `ur_signals.parquet` (~125 Hz) and `ur_signals_10hz.parquet` (10 Hz) cover **disjoint** UR3 episode subsets — no episode appears in both files. Together they span 7,060 distinct UR3 episodes; `ur_screwdriver_signals.parquet` adds another 1,240 screwdriver-task episodes. Each row in `episodes.parquet` corresponds to exactly one signal table.
100
+
101
+ ## Knowledge graph
102
+
103
+ `knowledge_graph/knowledge_graph.json` is the structured "world model" that grounds FactoryBench Q&A items: machine and gripper capability tables, the task/event vocabulary, the fault root-cause catalogue, the observable-anomaly catalogue, the error→protocol mapping used to derive ground-truth answers for L4 troubleshooting items, an anomaly severity ranking, and the relevance specs that drive fault-aware sub-series sampling. See [`knowledge_graph/SCHEMA.md`](https://huggingface.co/datasets/Forgis/FactoryBench/blob/main/knowledge_graph/SCHEMA.md) for field-level documentation.
104
+
105
+ ```python
106
+ import json, urllib.request
107
+ url = "https://huggingface.co/datasets/Forgis/FactoryBench/resolve/main/knowledge_graph/knowledge_graph.json"
108
+ kg = json.loads(urllib.request.urlopen(url).read())
109
+
110
+ # Machine spec lookup
111
+ machines_by_id = {m["machine_id"]: m for m in kg["machines"]}
112
+
113
+ # Error → operator protocol (the ground truth for L4 troubleshooting answers)
114
+ protocol_for = {e["root_cause"]: e["ur3_protocol"] for e in kg["root_cause_error_mapping"]}
115
+ print(protocol_for["collision_rigid_object"])
116
  ```
117
 
118
  ## Q&A pair counts
119
 
120
+ | Level | Full pool | Lite |
121
+ |-------|-----------|-------|
122
+ | L1 | 15,321 | 572 |
123
+ | L2 | 40,226 | 1,430 |
124
+ | L3 | 2,939 | 712 |
125
+ | L4 | 11,205 | 286 |
126
+ | **Total** | **69,691** | **3,000** |
127
+
128
+ The public release is a single undivided pool per level. The `train` split name
129
+ in the config block is the Hugging Face default for a single-file config, not a
130
+ semantic split. A private 15% slice is held back for contamination checks and is
131
+ not counted here.
132
 
133
  ## Q&A fields
134
 
135
+ Each line in `factorybench_qa/level_*.jsonl` (and `factorybench_lite/level_*.jsonl`) is a single Q&A item:
136
 
137
  | Field | Description |
138
  |-------|-------------|
 
154
  ```python
155
  from datasets import load_dataset
156
 
157
+ # Load one level of the full pool
158
+ ds = load_dataset("Forgis/FactoryBench", "level_1", split="train")
159
+
160
+ # Load one level of the balanced Lite subset
161
+ lite = load_dataset("Forgis/FactoryBench", "lite_level_1", split="train")
162
+
163
+ # Or address a file directly
164
  ds = load_dataset(
165
  "Forgis/FactoryBench",
166
+ data_files="factorybench_qa/level_1.jsonl",
167
  split="train",
168
  )
169
+
170
+ # Load underlying telemetry
171
+ import pandas as pd
172
+ root = "hf://datasets/Forgis/FactoryBench/factorywave"
173
+ episodes = pd.read_parquet(f"{root}/episodes.parquet")
174
+ ur_125 = pd.read_parquet(f"{root}/ur_signals.parquet")
175
+ ur_10 = pd.read_parquet(f"{root}/ur_signals_10hz.parquet") # different episodes
176
+ kuka = pd.read_parquet(f"{root}/kuka_signals.parquet")
177
+ screw = pd.read_parquet(f"{root}/ur_screwdriver_signals.parquet")
178
+
179
+ # Load the combined knowledge graph (machines, faults, error→protocol, ...)
180
+ import json, urllib.request
181
+ kg = json.loads(urllib.request.urlopen(
182
+ "https://huggingface.co/datasets/Forgis/FactoryBench/resolve/main/knowledge_graph/knowledge_graph.json"
183
+ ).read())
184
  ```
185
 
186
  ## Citation
 
188
  If you use FactoryBench, please cite the dataset and the two upstream open-source datasets it incorporates (AURSAD and voraus-AD):
189
 
190
  ```bibtex
191
+ @article{merzouki2026factorybench,
192
+ title = {FactoryBench: Evaluating Industrial Machine Understanding},
193
+ author = {Merzouki, Yanis and Izquierdo, Coral and Ignuta-Ciuncanu, Matei
194
+ and G\'omez-Bracamonte, Marcos and Maggioni, Riccardo and Lombardi,
195
+ Alessandro and Mazzoleni, Camilla and Martelli, Federico and
196
+ G\"unther, Bal\'azs and Petersen, Jonas and Petersen, Philipp},
197
+ journal = {arXiv preprint arXiv:2605.07675},
198
+ year = {2026}
199
  }
200
 
201
  @article{leporowski2022aursad,
202
  title = {{AURSAD}: Universal Robot Screwdriving Anomaly Detection Dataset},
203
+ author = {Leporowski, B{\l}a{\.z}ej and Tola, Daniella and Hansen, Casper and Iosifidis, Alexandros},
204
+ journal = {arXiv preprint arXiv:2102.01409},
205
+ year = {2021}
206
  }
207
 
208
  @misc{brockmann2024vorausad,
209
  title = {voraus-{AD}: A New Dataset for Anomaly Detection in Robot Applications},
210
  author = {Brockmann, Jan Thie{\ss} and Rudolph, Marco and Rosenhahn, Bodo and Wandt, Bastian},
211
  year = {2024},
212
+ eprint = {2311.04765},
213
  archivePrefix = {arXiv},
214
  primaryClass = {cs.RO}
215
  }
 
225
 
226
  ## License
227
 
228
+ Released under the [Creative Commons Attribution 4.0 International (CC BY 4.0)](https://creativecommons.org/licenses/by/4.0/) license. You may share and adapt the dataset for any purpose, including commercial, with appropriate attribution to the FactoryBench authors. The accompanying generator source code, evaluation scripts, LLM-as-judge prompts, and tooling (linked from the paper) are released separately under the MIT License.
factorybench_lite/level_1.jsonl ADDED
The diff for this file is too large to render. See raw diff
 
factorybench_lite/level_2.jsonl ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:6ef0ec3817a15df7b2a9927f046b3b01c8ed3935f653ee2eb98f0bebc70a4279
3
+ size 29363133
factorybench_lite/level_3.jsonl ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:fe357ea54f46aaaaa1045877ac688b2437440eba80bc30404a819b079a0e530e
3
+ size 20657794
factorybench_lite/level_4.jsonl ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:67b67a17934dafdaf31169c99806ed912d7e01ea6d2de82e376ebb1b9f4ad724
3
+ size 11346278
knowledge_graph/SCHEMA.md ADDED
@@ -0,0 +1,147 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # FactoryBench Knowledge Graph — Schema
2
+
3
+ `knowledge_graph.json` bundles the structured "world model" that grounds FactoryBench Q&A items: machine and gripper capabilities, the task/event vocabulary, the fault catalogue, and the operator-facing remediation protocols. The prompt builders and answer-derivation code consult these tables to (a) inject relevant context into prompts and (b) generate ground-truth answers for L4 troubleshooting items.
4
+
5
+ The file is a single JSON object with a top-level `$schema_version` and a `description`, plus the ten sections below. Every section value is either a `list` of records or a `dict`.
6
+
7
+ ## machines (list)
8
+
9
+ Robot arms used as test platforms. One entry per physical machine class.
10
+
11
+ | Field | Type | Notes |
12
+ |-------|------|-------|
13
+ | `machine_id` | int | Stable id referenced by `datasets[*].machine_id`. |
14
+ | `machine_model` | str | Vendor model code (e.g. `"UR3e"`). |
15
+ | `manufacturer`, `series` | str | Vendor metadata. |
16
+ | `machine_type` | str | E.g. `"collaborative robot"`. |
17
+ | `weight_kg`, `payload_kg`, `degrees_of_freedom` | num | Physical specs. |
18
+ | `joint_rotation` | list[num] | Per-joint rotation limits. |
19
+ | `typical_applications` | list[str] | Free-text capability tags. |
20
+ | `control_interfaces` | list[str] | E.g. `["UrScript", "PolyScope"]`. |
21
+ | `joint_speed_limits`, `rated_current_per_joint` | list[num] | Per-joint maxima. |
22
+ | `safety_modes`, `joint_modes`, `robot_modes`, `runtime_states` | list[obj] | Enum tables (id ↔ human-readable label). |
23
+ | `used_in_paper_for` | list[str] | Datasets/tasks this machine appears in. |
24
+
25
+ ## grippers (list)
26
+
27
+ End-effectors. One entry per gripper class.
28
+
29
+ | Field | Type | Notes |
30
+ |-------|------|-------|
31
+ | `gripper_id` | int | Stable id referenced by `datasets[*].gripper_id`. |
32
+ | `gripper_model`, `manufacturer`, `gripper_type`, `actuation` | str | Vendor metadata. |
33
+ | `finger_count`, `weight_kg`, `payload_kg`, `payload_kg_form_fit` | num | Physical specs. |
34
+ | `grip_force_range_N`, `torque_range_Nm`, `max_stroke_mm`, `opening_range_mm` | obj/list | Force/motion ranges. |
35
+
36
+ ## datasets (list)
37
+
38
+ The four episode collections that make up FactoryWave. One entry per (machine, gripper, task) combination.
39
+
40
+ | Field | Type | Notes |
41
+ |-------|------|-------|
42
+ | `dataset_id` | str | E.g. `"factorywave"`, `"aursad"`, `"vorausad"`. |
43
+ | `name`, `description` | str | Human-facing label. |
44
+ | `machine_id`, `gripper_id` | int | FK into `machines` / `grippers`. |
45
+ | `task_id` | int | FK into `tasks`. |
46
+ | `source` | str | URL or citation. |
47
+ | `license` | str | |
48
+
49
+ ## tasks (list)
50
+
51
+ Task-level vocabulary: each task is a sequence of named phases the robot moves through.
52
+
53
+ | Field | Type | Notes |
54
+ |-------|------|-------|
55
+ | `id` | int | Stable id referenced by `datasets[*].task_id` and used in `relevance_specs`. |
56
+ | `name`, `description` | str | E.g. `"pick_and_place"`. |
57
+ | `phases` | list[obj] | Ordered phase descriptors (id, name, intent). Used by phase-gated relevance sampling. |
58
+
59
+ ## events (list)
60
+
61
+ Atomic events that can occur during a task (e.g. collision triggers, gripper transitions, screwdriver phase changes).
62
+
63
+ | Field | Type | Notes |
64
+ |-------|------|-------|
65
+ | `id` | int | Stable event id encoded in episode rows under `event`. |
66
+ | `name`, `description` | str | |
67
+ | `tasks` | list[int] | Tasks where this event can occur. |
68
+ | `variables` | list[obj] | Per-event observable variables and their ranges. |
69
+
70
+ ## root_causes (list)
71
+
72
+ Catalogue of injectable fault mechanisms. The "physics" side of a fault.
73
+
74
+ | Field | Type | Notes |
75
+ |-------|------|-------|
76
+ | `fault_id` | int | Stable id used in episode-level `fault_label`. Also the FK from `root_cause_error_mapping`. |
77
+ | `task` | str | Which task this fault applies to. |
78
+ | `root_cause` | str | Snake-case identifier (e.g. `"collision_rigid_object"`). FK from `root_cause_error_mapping[*].root_cause`. |
79
+ | `description` | str | Plain-language explanation. |
80
+ | `severity_levels` | list[obj] | Mild / moderate / severe variants and their parameters. |
81
+ | `injectable` | bool | Whether the fault was actively injected (vs. passively observed). |
82
+ | `possible_anomalies` | list[str] | FK into `anomalies[*].anomaly_name`. |
83
+ | `simulation_procedure` | str | How to reproduce the fault. |
84
+ | `datasets` | list[str] | FK into `datasets[*].dataset_id`. |
85
+
86
+ ## anomalies (list)
87
+
88
+ Catalogue of *observable* symptoms (the "phenomenology" side of a fault).
89
+
90
+ | Field | Type | Notes |
91
+ |-------|------|-------|
92
+ | `anomaly_name` | str | Snake-case identifier (e.g. `"sudden_torque_spike"`). |
93
+ | `description` | str | What the anomaly looks like in the data. |
94
+ | `relevant_features_from_schema` | list[str] | Which channels (`feedback_speed_*`, `effort_target_torque_*`, etc.) the anomaly manifests on. |
95
+
96
+ A single fault can manifest as multiple anomalies, and a single anomaly can be caused by multiple faults — the join is via `root_causes[*].possible_anomalies`.
97
+
98
+ ## root_cause_error_mapping (list)
99
+
100
+ The error-to-protocol table. Maps each `fault_id` / `root_cause` to the UR3 controller error it raises (when any) and the operator-facing remediation protocol.
101
+
102
+ | Field | Type | Notes |
103
+ |-------|------|-------|
104
+ | `fault_id` | int | FK into `root_causes`. |
105
+ | `root_cause` | str | Mirror of `root_causes[*].root_cause` (denormalised for direct lookup). |
106
+ | `ur3_error_code` | str / null | UR3 controller error code (e.g. `"C 39 A 1"`), `null` for software-only faults. |
107
+ | `ur3_error_name` | str / null | Human-readable error name as shown on the teach pendant. |
108
+ | `ur3_description` | str / null | What the controller reports to the operator. |
109
+ | `ur3_protocol` | str | Step-by-step remediation procedure. **This is the ground-truth answer for L4 troubleshooting items.** |
110
+
111
+ ## anomaly_ranking (dict)
112
+
113
+ Severity ordering of anomalies, used by the rubric scorer.
114
+
115
+ | Field | Type | Notes |
116
+ |-------|------|-------|
117
+ | `ranking_least_to_most_severe` | list[str] | Anomaly names in ascending severity. |
118
+
119
+ ## relevance_specs (dict)
120
+
121
+ Drives fault-aware sub-series sampling — i.e. picking a window where the fault's signature is observable rather than uniformly random. See `src/question_generation/utils/relevance.py`.
122
+
123
+ | Field | Type | Notes |
124
+ |-------|------|-------|
125
+ | `version` | str | Spec version. |
126
+ | `description` | str | Human-facing summary. |
127
+ | `locality_definitions` | dict | The four sampling localities: `global` (uniform), `event` (window must contain transient), `phase_gated` (window must overlap target task phase), `cumulative` (length-based with optional phase gate). |
128
+ | `defaults` | dict | Per-locality default parameters merged into each spec. |
129
+ | `specs` | dict | Per-`fault_id` overrides. Keys are stringified fault ids; values declare locality, target phases, min window length, etc. |
130
+
131
+ Set `FB_RELEVANCE=0` in the environment to bypass and fall back to uniform sampling.
132
+
133
+ ## Loading
134
+
135
+ ```python
136
+ import json, urllib.request
137
+
138
+ URL = "https://huggingface.co/datasets/Forgis/FactoryBench/resolve/main/knowledge_graph/knowledge_graph.json"
139
+ kg = json.loads(urllib.request.urlopen(URL).read())
140
+
141
+ # Machine spec lookup by id
142
+ machines_by_id = {m["machine_id"]: m for m in kg["machines"]}
143
+
144
+ # Error→protocol lookup
145
+ protocol_for = {e["root_cause"]: e["ur3_protocol"] for e in kg["root_cause_error_mapping"]}
146
+ print(protocol_for["collision_rigid_object"])
147
+ ```
knowledge_graph/knowledge_graph.json ADDED
The diff for this file is too large to render. See raw diff