TransitLM: A Large-Scale Dataset and Benchmark for Map-Free Transit Route Generation

TransitLM is a large-scale dataset of over 13 million transit route planning records from four Chinese cities (Beijing, Shanghai, Shenzhen, Chengdu), covering 120,845 stations and 13,666 lines.

It is released as a continual pre-training corpus and benchmark data for three tasks: optimal route generation, preference-aware planning, and multi-route generation, evaluated with ten metrics across five categories.

Experiments show that an LLM trained on TransitLM produces structurally valid routes with high accuracy, implicitly grounds GPS coordinates to appropriate stations without explicit mapping, and generalizes across planning objectives with a single jointly trained model.

The dataset and code are publicly available on Hugging Face and GitHub, respectively.

The work demonstrates that transit route planning can be learned entirely from data, enabling end-to-end, map-free route generation directly from origin-destination information, bypassing traditional map-based routing engines.

The dataset includes route plans, station information, connectivity, and line details, with a benchmark of 10,000 test samples per task.

Results show that domain-specific models outperform general-purpose LLMs, which struggle with hallucinated stations and disconnected routes, and that continual pre-training is crucial for spatial grounding, as GPS-only input degrades general LLMs but not the trained models.

The joint model achieves up to 97.9% connectivity and 73.7% route exact match on optimal route generation, with minimal degradation under GPS-only input.

The dataset covers static route structures from a single platform, with future work planned for broader geographies and real-time dynamics.