---
license: apache-2.0
pretty_name: MetroLLM-Bench
language:
- en
task_categories:
- text-generation
tags:
- benchmark
- tool-calling
- agents
- transit
- kiosk
- llm-evaluation
size_categories:
- n<1K
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*.jsonl
  - split: holdout
    path: data/holdout-*.jsonl
---

# MetroLLM-Bench

The 955 cases of **MetroLLM-Bench**, a benchmark for language models as the policy layer of a
transit kiosk. The model receives kiosk events, calls structured tools (route planner, fare
calculator, station info, disruption feed, knowledge base) and submits a terminal state: outcome,
fare quote where applicable, kiosk action.

Paper: [MetroLLM-Bench: Evaluating Language Models as Transit Kiosk Runtimes](https://arxiv.org/abs/2609.10016)
(arXiv:2609.10016). Code, harness and ground-truth generator:
[github.com/continker/metrollm-bench](https://github.com/continker/metrollm-bench).
Fine-tuned students: the [MetroLLM-Bench v24 collection](https://huggingface.co/collections/continker/metrollm-bench-v24-6a35b586a11068e1b1ba3d47).
Live HF Space demo: [remcohendriks/metrollm](https://huggingface.co/spaces/remcohendriks/metrollm).

Mirror of `cases/*_cases.json` at tag `paper-v1.2` of the repository; the repository is the source of
truth. Regenerated by `scripts/build_hf_dataset.py`.

## Case details

One row per case. Each row holds the kiosk event sequence (station selected, passenger count
changed, disruption update, free-text input), the system context (current time, active
disruptions, framebook) and the ground truth used for scoring (expected route, fare, outcome,
kiosk action, and category-specific expectations such as traps or policies).

Scoring requires the harness in the repository: the mock tool server backed by the six system
datasets, the runner, and the 22-component scorer with its language-model judge. See
`REPRODUCING.md` there.

## Systems

| System | `system` | Cases | Train | Held-out |
|---|---|---:|---:|---:|
| MARTA (Atlanta) | `marta` | 156 | 117 | 39 |
| Doha Metro | `doha` | 156 | 117 | 39 |
| BART (San Francisco Bay Area) | `bart` | 157 | 118 | 39 |
| Taipei MRT | `taipei` | 167 | 125 | 42 |
| CTA L (Chicago) | `cta` | 157 | 118 | 39 |
| Beijing Subway | `beijing` | 162 | 122 | 40 |

## Categories

| `category` | Name | Cases |
|---|---|---:|
| A | Routing | 121 |
| B | Fare calculation | 93 |
| C | Disruptions | 124 |
| D | Accessibility | 91 |
| E | Cultural and multilingual | 42 |
| F | Policy change | 94 |
| G | Multi-turn | 90 |
| H | Adversarial | 90 |
| I | Temporal reasoning | 90 |
| J | Tool hallucination | 90 |
| K | Compound stress | 30 |

## Splits

`train` (717 cases) and `holdout` (238 cases) form the paper's stratified 75/25
case-level partition (seed 42, stratified by system). Training data for
the students came from `train` only; all held-out numbers in the paper are computed on `holdout`.

## Row schema

| Column | Type | Content |
|---|---|---|
| `id` | string | case id, e.g. `MARTA-A-001` |
| `system` | string | one of the six system keys |
| `category` | string | `A` to `K` |
| `difficulty` | string | `easy`, `medium`, `hard` |
| `interaction_mode` | string | `structured`, `freetext`, `multi_turn`, `adversarial`, `hallucination_probe`, `compound` |
| `title` | string | short scenario title |
| `partition` | string | `train` or `holdout` (same as the split) |
| `expected_outcome` | string | one of `route_and_fare_ready`, `advisory_only`, `service_unavailable`, `request_declined`, `policy_answer_only` |
| `expected_kiosk_action` | string | the kiosk action the terminal state must carry |
| `fare_total` | float or null | expected fare; null where no fare applies |
| `fare_currency` | string or null | `USD`, `QAR`, `TWD`, `CNY` |
| `events` | string (JSON) | the kiosk event sequence, verbatim |
| `system_context` | string (JSON) | current time, active disruptions, framebook, feature toggles |
| `ground_truth` | string (JSON) | the complete ground truth, all keys |
| `scoring` | string (JSON) | point weights of the scoring components that apply to this case |
| `tolerances` | string (JSON) | numeric tolerances used by the scorer |
| `extras` | string (JSON) | category-specific fields (scenario, trap, policy, temporal, accessibility, cultural ids) |

The JSON columns hold the original case objects verbatim; `json.loads` returns the same objects as
`cases/<system>_cases.json`.

## Loading

```python
from datasets import load_dataset
import json

ds = load_dataset("continker/metrollm-bench")
row = ds["holdout"][0]
events = json.loads(row["events"])
ground_truth = json.loads(row["ground_truth"])
```

## Citation

```bibtex
@techreport{hendriks2026metrollm,
  title       = {MetroLLM-Bench: Evaluating Language Models as Transit Kiosk Runtimes},
  author      = {Hendriks, Remco},
  institution = {Continker},
  type        = {Technical report},
  number      = {v1.2},
  year        = {2026},
  month       = {9},
  doi         = {10.5281/zenodo.21893944},
  eprint      = {2609.10016},
  archiveprefix = {arXiv},
  primaryclass = {cs.LG},
  url         = {https://arxiv.org/abs/2609.10016}
}
```

## License

Apache License 2.0, as the code, ground truth and framebooks in the repository.