umarfarookm commited on
Commit
9a3bbad
·
verified ·
1 Parent(s): e9ac851

Add dataset card

Browse files
Files changed (1) hide show
  1. README.md +155 -30
README.md CHANGED
@@ -1,32 +1,157 @@
1
  ---
2
- dataset_info:
3
- features:
4
- - name: instruction
5
- dtype: string
6
- - name: response
7
- dtype: string
8
- - name: category
9
- dtype: string
10
- - name: template_id
11
- dtype: string
12
- - name: feed_id
13
- dtype: string
14
- - name: provider
15
- dtype: string
16
- splits:
17
- - name: train
18
- num_bytes: 842262
19
- num_examples: 2971
20
- - name: test
21
- num_bytes: 92757
22
- num_examples: 335
23
- download_size: 240663
24
- dataset_size: 935019
25
- configs:
26
- - config_name: default
27
- data_files:
28
- - split: train
29
- path: data/train-*
30
- - split: test
31
- path: data/test-*
32
  ---
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
  ---
2
+ license: apache-2.0
3
+ task_categories:
4
+ - text-generation
5
+ - question-answering
6
+ language:
7
+ - en
8
+ tags:
9
+ - transit
10
+ - gtfs
11
+ - transportation
12
+ - public-transit
13
+ - instruction-tuning
14
+ - synthetic
15
+ size_categories:
16
+ - 1K<n<10K
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
17
  ---
18
+
19
+ # UmarTransit-Instruct-3k
20
+
21
+ A synthetic instruction-tuning dataset for **public transit systems** and **GTFS (General Transit Feed Specification)**, containing 3,306 question-answer pairs generated from 15 real-world open GTFS feeds across 10 countries.
22
+
23
+ Built to train domain-specific language models like [UmarTransit-1B](https://huggingface.co/umarfarookm/UmarTransit-1B).
24
+
25
+ > **Data Disclaimer:** This dataset was generated **exclusively from publicly available, open-source GTFS feeds** published by transit agencies for public use via the [Mobility Database](https://mobilitydatabase.org/). **No private, proprietary, or NDA-protected data** from any client, employer, or organization was used.
26
+
27
+ ## Dataset Details
28
+
29
+ | Property | Value |
30
+ |----------|-------|
31
+ | **Total pairs** | 3,306 |
32
+ | **Training split** | 2,971 (90%) |
33
+ | **Test split** | 335 (10%) |
34
+ | **Categories** | 8 task categories |
35
+ | **Templates** | 45 question templates |
36
+ | **GTFS feeds** | 15 feeds from 10 countries |
37
+ | **Format** | JSONL (one JSON object per line) |
38
+ | **Language** | English |
39
+ | **License** | Apache 2.0 |
40
+
41
+ ## Usage
42
+
43
+ ```python
44
+ from datasets import load_dataset
45
+
46
+ dataset = load_dataset("umarfarookm/UmarTransit-Instruct-3k")
47
+
48
+ # Access splits
49
+ train = dataset["train"]
50
+ test = dataset["test"]
51
+
52
+ # Example
53
+ print(train[0]["instruction"])
54
+ print(train[0]["response"])
55
+ ```
56
+
57
+ ## Data Format
58
+
59
+ Each record contains:
60
+
61
+ ```json
62
+ {
63
+ "instruction": "How many routes does Chicago Transit Authority (CTA) have?",
64
+ "response": "Chicago Transit Authority (CTA) operates 133 routes. 4 are Tram/Streetcar/Light rail routes...",
65
+ "category": "agency_overview",
66
+ "template_id": "agency_route_count_v1",
67
+ "feed_id": "389",
68
+ "provider": "Chicago Transit Authority (CTA)"
69
+ }
70
+ ```
71
+
72
+ | Field | Description |
73
+ |-------|-------------|
74
+ | `instruction` | The user question |
75
+ | `response` | The expected answer |
76
+ | `category` | Task category (1 of 8) |
77
+ | `template_id` | Which template generated this pair |
78
+ | `feed_id` | Source GTFS feed ID from Mobility Database |
79
+ | `provider` | Transit agency name |
80
+
81
+ ## Task Categories
82
+
83
+ | Category | Count | Description |
84
+ |----------|-------|-------------|
85
+ | agency_overview | 1,075 | Agency transit modes, route counts, timezones |
86
+ | stop_info | 911 | Stop locations, coordinates, accessibility |
87
+ | schedule | 636 | Trip schedules, departure/arrival times |
88
+ | route_info | 457 | Route descriptions, types, trip counts |
89
+ | transfer | 161 | Transfer connections, types, wait times |
90
+ | network_stats | 30 | Aggregate network statistics |
91
+ | gtfs_knowledge | 22 | GTFS specification concepts and definitions |
92
+ | comparative | 14 | Cross-agency comparisons |
93
+
94
+ ## Source GTFS Feeds
95
+
96
+ All feeds are publicly available through the [Mobility Database](https://mobilitydatabase.org/).
97
+
98
+ | Country | City/Region | Agency | Feed ID |
99
+ |---------|-------------|--------|---------|
100
+ | US | Los Angeles | LA Metro | 29 |
101
+ | US | Chicago | CTA | 389 |
102
+ | US | Boston | MBTA | 437 |
103
+ | US | Phoenix | Valley Metro | 1086 |
104
+ | US | Austin | Capital Metro | 1029 |
105
+ | US | Portland | TriMet | 1077 |
106
+ | Canada | Toronto | TTC | 247 |
107
+ | Germany | Berlin | VBB | 782 |
108
+ | France | Paris | Ile-de-France Mobilites | 865 |
109
+ | Netherlands | National | OVapi | 1292 |
110
+ | Belgium | National | NMBS/SNCB | 732 |
111
+ | Finland | Helsinki | HSL | 686 |
112
+ | Denmark | National | Rejseplanen | 150 |
113
+ | Australia | Perth | Transperth | 1026 |
114
+ | New Zealand | Auckland | Auckland Transport | 147 |
115
+
116
+ ## Generation Process
117
+
118
+ 1. **Download** 15 open GTFS feeds from the Mobility Database
119
+ 2. **Clean** raw CSV data into normalized Parquet format
120
+ 3. **Extract** feed statistics (routes, stops, trips, transfers, schedules)
121
+ 4. **Generate** Q&A pairs using 45 templates across 8 categories
122
+ 5. **Validate** all pairs for format, content quality, and factual accuracy
123
+ 6. **Split** into train/test (90/10, stratified by category)
124
+
125
+ All scripts are open-source: [github.com/umarfarookm/transit-foundation-model](https://github.com/umarfarookm/transit-foundation-model)
126
+
127
+ ## Quality Validation
128
+
129
+ - **Format errors:** 0 / 3,306
130
+ - **Duplicate instructions:** 0
131
+ - **Factual accuracy:** 100% (275 spot-checks against source data)
132
+ - **Average instruction length:** 66 characters
133
+ - **Average response length:** 136 characters
134
+
135
+ ## Trained Model
136
+
137
+ This dataset was used to train [UmarTransit-1B](https://huggingface.co/umarfarookm/UmarTransit-1B), which shows a **+74% improvement** over the base model (Qwen2.5-1.5B-Instruct) on a 193-question benchmark evaluation.
138
+
139
+ ## Limitations
140
+
141
+ - **English only** — no multilingual coverage
142
+ - **Static schedules** — no real-time or delay data
143
+ - **Template-based** — all Q&A pairs follow fixed templates, limiting response diversity
144
+ - **15 feeds** — does not cover all transit agencies worldwide
145
+ - **Small scale** — 3,306 pairs is modest compared to general instruction datasets
146
+
147
+ ## Citation
148
+
149
+ ```bibtex
150
+ @dataset{umartransit_instruct_3k,
151
+ author = {Umar Farook M},
152
+ title = {UmarTransit-Instruct-3k: Transit and GTFS Instruction Dataset},
153
+ year = {2026},
154
+ publisher = {Hugging Face},
155
+ url = {https://huggingface.co/datasets/umarfarookm/UmarTransit-Instruct-3k}
156
+ }
157
+ ```