README.md

metadata

license: apache-2.0
task_categories:
  - text-generation
  - question-answering
language:
  - en
tags:
  - transit
  - gtfs
  - transportation
  - public-transit
  - instruction-tuning
  - synthetic
size_categories:
  - 1K<n<10K
dataset_info:
  features:
    - name: instruction
      dtype: string
    - name: response
      dtype: string
    - name: category
      dtype: string
    - name: template_id
      dtype: string
    - name: feed_id
      dtype: string
    - name: provider
      dtype: string
  splits:
    - name: train
      num_bytes: 904730
      num_examples: 3154
    - name: test
      num_bytes: 97857
      num_examples: 347
  download_size: 274812
  dataset_size: 1002587
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
      - split: test
        path: data/test-*

UmarTransit-Instruct-3k

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.

Built to train domain-specific language models like UmarTransit-1B.

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. No private, proprietary, or NDA-protected data from any client, employer, or organization was used.

Dataset Details

Property	Value
Total pairs	3,306
Training split	2,971 (90%)
Test split	335 (10%)
Categories	8 task categories
Templates	45 question templates
GTFS feeds	15 feeds from 10 countries
Format	JSONL (one JSON object per line)
Language	English
License	Apache 2.0

Usage

from datasets import load_dataset

dataset = load_dataset("umarfarookm/UmarTransit-Instruct-3k")

# Access splits
train = dataset["train"]
test = dataset["test"]

# Example
print(train[0]["instruction"])
print(train[0]["response"])

Data Format

Each record contains:

{
  "instruction": "How many routes does Chicago Transit Authority (CTA) have?",
  "response": "Chicago Transit Authority (CTA) operates 133 routes. 4 are Tram/Streetcar/Light rail routes...",
  "category": "agency_overview",
  "template_id": "agency_route_count_v1",
  "feed_id": "389",
  "provider": "Chicago Transit Authority (CTA)"
}

Field	Description
instruction	The user question
response	The expected answer
category	Task category (1 of 8)
template_id	Which template generated this pair
feed_id	Source GTFS feed ID from Mobility Database
provider	Transit agency name

Task Categories

Category	Count	Description
agency_overview	1,075	Agency transit modes, route counts, timezones
stop_info	911	Stop locations, coordinates, accessibility
schedule	636	Trip schedules, departure/arrival times
route_info	457	Route descriptions, types, trip counts
transfer	161	Transfer connections, types, wait times
network_stats	30	Aggregate network statistics
gtfs_knowledge	22	GTFS specification concepts and definitions
comparative	14	Cross-agency comparisons

Source GTFS Feeds

All feeds are publicly available through the Mobility Database.

Country	City/Region	Agency	Feed ID
US	Los Angeles	LA Metro	29
US	Chicago	CTA	389
US	Boston	MBTA	437
US	Phoenix	Valley Metro	1086
US	Austin	Capital Metro	1029
US	Portland	TriMet	1077
Canada	Toronto	TTC	247
Germany	Berlin	VBB	782
France	Paris	Ile-de-France Mobilites	865
Netherlands	National	OVapi	1292
Belgium	National	NMBS/SNCB	732
Finland	Helsinki	HSL	686
Denmark	National	Rejseplanen	150
Australia	Perth	Transperth	1026
New Zealand	Auckland	Auckland Transport	147

Generation Process

1. Download 15 open GTFS feeds from the Mobility Database
2. Clean raw CSV data into normalized Parquet format
3. Extract feed statistics (routes, stops, trips, transfers, schedules)
4. Generate Q&A pairs using 45 templates across 8 categories
5. Validate all pairs for format, content quality, and factual accuracy
6. Split into train/test (90/10, stratified by category)

All scripts are open-source: github.com/umarfarookm/transit-foundation-model

Quality Validation

• Format errors: 0 / 3,306
• Duplicate instructions: 0
• Factual accuracy: 100% (275 spot-checks against source data)
• Average instruction length: 66 characters
• Average response length: 136 characters

Trained Model

This dataset was used to train UmarTransit-1B, which shows a +74% improvement over the base model (Qwen2.5-1.5B-Instruct) on a 193-question benchmark evaluation.

Limitations

• English only — no multilingual coverage
• Static schedules — no real-time or delay data
• Template-based — all Q&A pairs follow fixed templates, limiting response diversity
• 15 feeds — does not cover all transit agencies worldwide
• Small scale — 3,306 pairs is modest compared to general instruction datasets

Citation

@dataset{umartransit_instruct_3k,
  author = {Umar Farook M},
  title = {UmarTransit-Instruct-3k: Transit and GTFS Instruction Dataset},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/umarfarookm/UmarTransit-Instruct-3k}
}