zhangdw/TurnBack-10pct

🧭 TurnBack 10% Stratified Sample

A deterministic route-reversal subset for lightweight geospatial cognition experiments

    

TurnBack 10% is a balanced, reproducible subset of the public TurnBack route corpus for route-level reverse-navigation generation and evaluation.

Quick Start · At a Glance · Files · Sampling · Citation

---

This is a 10% stratified sample of the public raw route corpus from bghjmn32/EMNLP2025_Turnback, not the full TurnBack benchmark. It is intended for lightweight route-reversal experiments, prompt iteration, and reproducibility checks before full-scale evaluation.

✨ Why This Dataset?

TurnBack evaluates route-level geospatial cognition: a model receives forward pedestrian navigation instructions and must generate reverse instructions that guide the pedestrian back to the original start. The reference geometry is the original route traversed backwards, and downstream evaluation should execute the generated reverse instructions with Path Builder before scoring recovered geometry.

This subset is designed for questions like:

• Can a model reverse pedestrian navigation instructions across cities and difficulty levels?
• How does route-reversal quality change from easy to hard routes?
• Can a geospatial evaluator run quickly on a balanced sample before full TurnBack?
• Can every sampled route be traced back to official raw instructions and route geometry?

📦 Dataset at a Glance

3,900 sampled routes	13 public city directories	3 difficulty levels	100 routes per stratum
1,300 easy routes	1,300 medium routes	1,300 hard routes	20260606 sampling seed

Covered cities

Auckland_New_Zealand · Cairo_Egypt · Cape_Town_South_Africa · Denver_Colorado_USA · London_UK · Mexico_City_Mexico · Munich_Germany · Paris_France · Rio_de_Janeiro_Brazil · Singapore_Singapore · Sydney_Australia · Tokyo_23_wards · Toronto_Canada

Evaluation shape

Component	Description
Input	Forward pedestrian navigation instructions from the official raw corpus.
Output	Model-generated reverse instructions back to the original start.
Reference	Original route geometry traversed backwards.
Recommended scorer	Execute generated instructions with Path Builder, then compare recovered route geometry.
Interaction pattern	One prompt / response per sampled route.

🚀 Quick Start

Load the JSONL dataset:

from datasets import load_dataset

ds = load_dataset(
    "json",
    data_files="https://huggingface.co/datasets/zhangdw/TurnBack-10pct/resolve/main/data/turnback_10pct.jsonl",
    split="train",
)

print(len(ds))
print(ds[0].keys())

Inspect one route-reversal prompt:

row = ds[0]
print(row["city"], row["difficulty"], row["route_id"])
print(row["instructions_text"][:1000])
print(row["prompt"][:1000])

Download the full subset, metadata, and regeneration script:

hf download zhangdw/TurnBack-10pct \
  --type dataset \
  --local-dir data/turnback-10pct

🗂️ Files

Path	Description
data/turnback_10pct.jsonl	One JSON object per sampled route.
metadata/counts.json	Sampling settings, city/difficulty counts, and validation summary.
metadata/sample_ids.json	Sampled route IDs for every city × difficulty cell.
scripts/build_turnback_10pct_dataset.py	Exact script used to regenerate this dataset from the official source.

Each JSONL record contains:

Field	Description
id, city, difficulty, route_id	Stable route identifiers and strata.
instructions_text	Forward navigation instructions from the official raw corpus.
route_geojson	Reference route geometry from the official raw corpus.
route_summary	Distance, duration, step count, coordinate count, bbox, and start/end coordinates.
prompt	A simple route-reversal prompt built from the forward instructions.
source_paths, source_urls	Provenance back to the official GitHub raw files.

🧭 Example Workflows

Count routes by city and difficulty

import pandas as pd

df = ds.to_pandas()
print(df.groupby(["city", "difficulty"]).size())
print(df.groupby("difficulty").size())

Read the sampling manifest

import json
from huggingface_hub import hf_hub_download

path = hf_hub_download(
    repo_id="zhangdw/TurnBack-10pct",
    repo_type="dataset",
    filename="metadata/counts.json",
)

with open(path, "r", encoding="utf-8") as f:
    counts = json.load(f)

print(counts["total_sampled_routes"])
print(counts["counts_by_difficulty"])

Build a simple model input

messages = [
    {
        "role": "system",
        "content": "You reverse pedestrian navigation instructions while preserving route geometry.",
    },
    {"role": "user", "content": row["prompt"]},
]

🧪 Sampling Policy

Setting	Value
Source repository	bghjmn32/EMNLP2025_Turnback
Source revision	main
Source root	36kroutes/
Sampling seed	20260606
Sampling fraction	0.1
Stratification	city × difficulty
Sample per cell	100 routes
Route ID universe per cell	0..999
Total sampled routes	3,900
Created at	2026-06-06T18:31:38.193396+00:00

The original paper describes a 36,000-route benchmark over 12 metropolitan areas. The currently published raw 36kroutes/ snapshot in the official GitHub repository contains an additional public city directory, Paris_France. The official public code can enumerate, load, and score that city, so this subset includes all 13 public city directories. To preserve a paper-style balanced layout, it samples from the canonical route-id universe 0..999 in each city × difficulty cell and excludes overflow route directories that appear in some raw-snapshot cells.

📊 Counts by Stratum

Every city contributes 300 routes: 100 easy, 100 medium, and 100 hard.

Full city × difficulty table

City	Easy	Medium	Hard	Total
Auckland_New_Zealand	100	100	100	300
Cairo_Egypt	100	100	100	300
Cape_Town_South_Africa	100	100	100	300
Denver_Colorado_USA	100	100	100	300
London_UK	100	100	100	300
Mexico_City_Mexico	100	100	100	300
Munich_Germany	100	100	100	300
Paris_France	100	100	100	300
Rio_de_Janeiro_Brazil	100	100	100	300
Singapore_Singapore	100	100	100	300
Sydney_Australia	100	100	100	300
Tokyo_23_wards	100	100	100	300
Toronto_Canada	100	100	100	300
Total	1,300	1,300	1,300	3,900

✅ Intended Use

This subset is useful for:

• quick route-reversal prompt and evaluator iteration;
• balanced city × difficulty smoke tests;
• geospatial instruction-following experiments;
• provenance-preserving route-level error analysis;
• validating code before running the full TurnBack corpus.

It is not a replacement for full TurnBack benchmark reporting.

⚖️ License

This subset follows the MIT license metadata of the upstream public TurnBack release. Check the official TurnBack repository and paper for source-specific requirements.

📚 Citation

If you use this dataset, cite the original TurnBack paper and official code/data release:

@inproceedings{luo2025turnback,
  title = {TurnBack: A Geospatial Route Cognition Benchmark for Large Language Models through Reverse Route},
  author = {Luo, Hongyi and Cheng, Qing and Matos, Daniel and Gadi, Hari Krishna and Zhang, Yanfeng and Liu, Lu and Wang, Yongliang and Zeller, Niclas and Cremers, Daniel and Meng, Liqiu},
  booktitle = {Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing},
  year = {2025}
}

You may also cite this deterministic subset when the 10% sampling policy matters for reproducibility:

@misc{turnback_10pct2026,
  title        = {TurnBack 10% Stratified Sample},
  author       = {Dawei Zhang},
  year         = {2026},
  howpublished = {Hugging Face Dataset},
  url          = {https://huggingface.co/datasets/zhangdw/TurnBack-10pct}
}

---

A compact route-reversal sample for balanced geospatial cognition experiments.