UmarTransit-1B: open-source GTFS domain LLM fine-tuned from Qwen2.5-1.5B

A developer has released UmarTransit-1B, an open-source language model specialized for public transit systems and GTFS data, built by fine-tuning the Qwen2.5-1.5B-Instruct base model.

The model was trained using QLoRA on 3,306 synthetic instruction-response pairs derived from 77M+ rows of cleaned GTFS feeds across 10 countries. It achieves a ROUGE-L score of 0.82 overall and is available in Safetensors and GGUF formats for local inference.

The project demonstrates that data quality outweighs quantity, showing that well-structured training pairs can achieve strong results on free hardware like a Colab T4.