Datasets:

Modalities:
Text
Formats:
parquet
Languages:
German
Tags:
legal
DOI:
License:
kapllan commited on
Commit
8434a80
·
verified ·
1 Parent(s): 84086ff

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +67 -12
README.md CHANGED
@@ -8,11 +8,11 @@ configs:
8
  path: "eval.parquet"
9
  - split: test
10
  path: "test.parquet"
11
- - split: complete
12
- path: "all.parquet"
 
13
  - split: metadata
14
  path: "project_summaries.parquet"
15
- default: true
16
 
17
  license: cc-by-4.0
18
  task_categories:
@@ -61,6 +61,7 @@ These annotations are based on a specific snapshot of the sustainability procure
61
  - `text_cleaned` (string): Normalized and cleaned text
62
  - `catalog` (string): Sector-specific guideline (e.g., "Road_Transport", "Print_Services", "ICT", "Food", etc.)
63
  - `area_of_action_id` (string): Specific criterion ID within the catalog (e.g., "HF-001", or "not present" for negative examples)
 
64
  - `procurement_criterion` (string): Legal context type (e.g., "technical_specification", "award_criterion", "selection_criterion", "terms_and_conditions", "evidence_requirement", or "not present")
65
  - `self_declaration` (string): Indicates if it's a self-declaration ("True"/"False" as string, or "not present")
66
  - `evidence` (string): Evidence requirements (or "not present")
@@ -68,16 +69,32 @@ These annotations are based on a specific snapshot of the sustainability procure
68
 
69
  ### Data Splits
70
 
71
- The dataset provides a **stratified random split**:
72
  - **50% Training**
73
- - **25% Evaluation**
74
  - **25% Test**
75
 
76
- Stratification ensures consistent distribution across:
77
- - `area_of_action_id` (criterion ID)
78
- - `procurement_criterion` (criterion type)
79
- - `pos_neg` (positive/negative label - internal mapping)
80
- - `award_criterion_weight` (award criteria weights)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
81
 
82
  ### Sector-Specific Guidelines (Catalogs)
83
 
@@ -92,6 +109,42 @@ The dataset includes sustainability criteria from multiple sectors:
92
  - Product Independent Criteria
93
  - And others
94
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
95
  ## Dataset Creation
96
 
97
  ### Source Data
@@ -110,7 +163,7 @@ The generation pipeline includes:
110
  2. **Negative Sampling:**
111
  - **Hard Negatives:** Sentences with high semantic similarity (cosine similarity) but no sustainability criteria
112
  - **Random Negatives:** Random sentences for standard administrative language
113
- 3. **Contextual Windowing:** Creates context windows (1 or 2 surrounding sentences) with label leakage prevention (available in `dataset_info.json` but excluded from primary splits by default)
114
  4. **Text Cleaning:** Normalization of line breaks, removal of excess whitespace (available in `text_cleaned`)
115
  5. **Deduplication:** Based on `project_id`, `simap_version`, `text_cleaned`, `filename`, `catalog`, and `area_of_action_id`
116
 
@@ -132,6 +185,7 @@ Annotation categories (`procurement_criterion`) include:
132
 
133
  - **NLP Engine:** spaCy (model: de_dep_news_trf) for German sentence segmentation
134
  - **Embeddings:** SentenceTransformers (all-MiniLM-L6-v2) for semantic search and hard negative mining
 
135
  - **Environment:** Python 3.10 via condavenv
136
 
137
  ## Considerations
@@ -145,10 +199,11 @@ This dataset supports:
145
 
146
  ### Limitations
147
 
148
- - **Language:** Dataset is German-only, specific to Swiss German administrative language
149
  - **Domain Specificity:** Trained on Swiss public procurement documents; may not generalize to other jurisdictions
150
  - **Context Dependency:** Some criteria require understanding of broader document context
151
  - **Class Imbalance:** Distribution between positive and negative examples should be monitored during training (mapped via `area_of_action_id` != "not present")
 
152
 
153
  ## Citation
154
 
 
8
  path: "eval.parquet"
9
  - split: test
10
  path: "test.parquet"
11
+ default: true
12
+ - config_name: metadata
13
+ data_files:
14
  - split: metadata
15
  path: "project_summaries.parquet"
 
16
 
17
  license: cc-by-4.0
18
  task_categories:
 
61
  - `text_cleaned` (string): Normalized and cleaned text
62
  - `catalog` (string): Sector-specific guideline (e.g., "Road_Transport", "Print_Services", "ICT", "Food", etc.)
63
  - `area_of_action_id` (string): Specific criterion ID within the catalog (e.g., "HF-001", or "not present" for negative examples)
64
+ - `sustainability_dimensions` (string): Sustainability dimensions of the annotated area of action as a bracketed list, e.g. "[ecologic, social]" ("[]" for negative examples or when no dimension applies)
65
  - `procurement_criterion` (string): Legal context type (e.g., "technical_specification", "award_criterion", "selection_criterion", "terms_and_conditions", "evidence_requirement", or "not present")
66
  - `self_declaration` (string): Indicates if it's a self-declaration ("True"/"False" as string, or "not present")
67
  - `evidence` (string): Evidence requirements (or "not present")
 
69
 
70
  ### Data Splits
71
 
72
+ The dataset provides a **coverage-aware, leakage-safe split**:
73
  - **50% Training**
74
+ - **25% Evaluation**
75
  - **25% Test**
76
 
77
+ Splitting happens at the *document-snippet* level: a snippet and all of its
78
+ annotation rows always land in the same split. Two seeded (and therefore
79
+ reproducible) phases then allocate the snippet groups:
80
+ 1. **Coverage** — every `(catalog, area_of_action_id)` pair of the positively
81
+ annotated snippets and every real `procurement_criterion` value is
82
+ processed rarest-first and placed into splits that do not contain it yet.
83
+ Pairs held by 3 or more snippets are guaranteed to appear in all three
84
+ splits; rarer pairs cover as many splits as possible and are flagged in the
85
+ `Split_Coverage` sheet of `statistics.xlsx`.
86
+ 2. **Balancing** — the remaining snippets go to the split where their labels
87
+ lag furthest behind the ideal 50/25/25 quota, so `procurement_criterion`
88
+ shares typically deviate by less than 1 % from the ideal proportions.
89
+
90
+ `statistics.xlsx` additionally reports how the annotations are distributed
91
+ across the splits at row level: the `Split_Areas_Of_Action`, `Split_Catalogs`
92
+ and `Split_Criteria` sheets each list every label with its total annotation
93
+ count, its count per split and the share of the label's annotations that each
94
+ split receives.
95
+
96
+ The sentinels "not present"/"not specified" never count as labels; negative
97
+ snippets carry no labels and are distributed purely by remaining split capacity.
98
 
99
  ### Sector-Specific Guidelines (Catalogs)
100
 
 
109
  - Product Independent Criteria
110
  - And others
111
 
112
+ ### Metadata Subset (`project_summaries`)
113
+
114
+ Besides the annotated text data, the repository publishes a per-project
115
+ metadata subset — `project_summaries.json` / `project_summaries.parquet`, one
116
+ row per procurement project (~56 rows). It documents *who annotated what* and
117
+ *which sector catalogs came into question*, so the text data can be put into
118
+ project context without exposing the annotation workflow's internal
119
+ bookkeeping. On Hugging Face the subset is published as its own config/split
120
+ pair — config `metadata`, split `metadata` (see the `configs:` block above) —
121
+ so the Dataset Viewer shows it in the split selector next to the `default`
122
+ config's `train`/`eval`/`test` splits, and
123
+ `load_dataset("IntelliProcure/SwissSPARK", "metadata", split="metadata")`
124
+ loads it directly:
125
+
126
+ | Field | Type | Description |
127
+ |---|---|---|
128
+ | `project_id` | int | SIMAP project ID |
129
+ | `simap_version` | string | SIMAP platform version (e.g. "simap") |
130
+ | `division` / `cpv_codes` | string | CPV division text / CPV codes of the project |
131
+ | `annotators` | list of strings | **Pseudonymized** annotator IDs (see below) |
132
+ | `candidate_catalogs` | list of strings | Catalogs whose workbook column is marked `match` — they came into question for this project |
133
+ | `annotated_catalogs` | list of strings | Catalogs marked `annotated` — they were actually annotated for this project |
134
+ | `documents` | list of strings | The project's documents (dataset filenames, joinable to the main subset's `filename`) |
135
+
136
+ Catalog names equal the `catalog` vocabulary of the default config (e.g.
137
+ `Food`, `Road_Transport`, `Product_Independent_Criteria`).
138
+
139
+ **Annotator pseudonymization:** annotator names are never published raw. For
140
+ each release, all annotators recorded in the source workbook are sorted
141
+ alphabetically and mapped onto `annotator_1` … `annotator_N`; the mapping is
142
+ identical in every row. The IDs are **release-scoped**: they are regenerated
143
+ on every rebuild and the numbering can shift between releases when the
144
+ annotator roster changes. Columns that may contain personal data ("Recommended
145
+ by", internal comments) are never exported, and generation fails hard if a
146
+ real annotator name is detected anywhere in the subset.
147
+
148
  ## Dataset Creation
149
 
150
  ### Source Data
 
163
  2. **Negative Sampling:**
164
  - **Hard Negatives:** Sentences with high semantic similarity (cosine similarity) but no sustainability criteria
165
  - **Random Negatives:** Random sentences for standard administrative language
166
+ 3. **Contextual Windowing:** Creates context windows (1 or 2 surrounding sentences) around each excerpt; negative candidates overlapping annotated excerpts are skipped during mining instead of being published with an in-context warning
167
  4. **Text Cleaning:** Normalization of line breaks, removal of excess whitespace (available in `text_cleaned`)
168
  5. **Deduplication:** Based on `project_id`, `simap_version`, `text_cleaned`, `filename`, `catalog`, and `area_of_action_id`
169
 
 
185
 
186
  - **NLP Engine:** spaCy (model: de_dep_news_trf) for German sentence segmentation
187
  - **Embeddings:** SentenceTransformers (all-MiniLM-L6-v2) for semantic search and hard negative mining
188
+ - **Language filter:** lingua (restricted to German/French/Italian/English) excludes non-German sentences from mined negative examples
189
  - **Environment:** Python 3.10 via condavenv
190
 
191
  ## Considerations
 
199
 
200
  ### Limitations
201
 
202
+ - **Language:** Dataset is German-only, specific to Swiss German administrative language (mined negative examples are filtered to German-only sentences via lingua; human-annotated positives are kept as annotated)
203
  - **Domain Specificity:** Trained on Swiss public procurement documents; may not generalize to other jurisdictions
204
  - **Context Dependency:** Some criteria require understanding of broader document context
205
  - **Class Imbalance:** Distribution between positive and negative examples should be monitored during training (mapped via `area_of_action_id` != "not present")
206
+ - **Release-Scoped Annotator Pseudonyms:** the `annotator_*` IDs of the `project_summaries` config are recomputed from the source workbook on every release and may change between releases; they identify a person only within one published version
207
 
208
  ## Citation
209