kxiaoqiangrexian/MTS-All

MTS-All

MTS-All is the data release for the EMNLP 2026 accepted paper **Reactivating Test-Time Scaling for Plane Geometry Problem Solving** (PDF).

Training and evaluation code is available in the ReTTS-PGPS repository.

The dataset contains multi-trace supervised fine-tuning data and test files for three plane geometry benchmarks:

• —PGPS9K-All
• —Geometry3K-All
• —GeoQA-All

Each training problem is represented with four reasoning traces:

• —Program: symbolic geometry program.
• —COT-program: chain-of-thought augmented symbolic program.
• —PAL: executable Python / SymPy program-aided reasoning trace.
• —COT-PAL: chain-of-thought augmented PAL trace.

The data is intended for training and evaluating multimodal geometry solvers under perception-augmented training, self-consistency, Standard MTE, and CG-MTE.

Directory Structure

.
├── train/
│   ├── pgps9k/
│   │   ├── Program/train/train.parquet
│   │   ├── COT-program/train/train.parquet
│   │   ├── PAL/train/train.parquet
│   │   └── COT-PAL/train/train.parquet
│   ├── geometry3k/
│   │   ├── Program/train/train.parquet
│   │   ├── COT-program/train/train.parquet
│   │   ├── PAL/train/train.parquet
│   │   └── COT-PAL/train/train.parquet
│   └── geoqa/
│       ├── Program/train/train.parquet
│       ├── COT-program/train/train.parquet
│       ├── PAL/train/train.parquet
│       └── COT-PAL/train/train.parquet
└── test/
    ├── pgps9k/
    │   ├── test.jsonl
    │   ├── test.parquet
    │   └── images/
    ├── geometry3k/
    │   ├── test.jsonl
    │   ├── test.parquet
    │   └── images/
    └── geoqa/
        ├── test.jsonl
        ├── test.parquet
        └── images/

Data Statistics

The following counts are computed from the uploaded files in this release.

The paper describes these as approximately 28K / 30K / 14K MTS-All training instances after multi-trace synthesis and filtering. The table above reports the exact counts of this Hugging Face release.

The test split contains 893 unique PGPS9K images, 519 unique Geometry3K images, and 754 unique GeoQA images. Multiple questions that share the same diagram reference the same image file.

File Format

Training Parquet

Each training parquet file contains:

• —prob_id: problem identifier.
• —image: image feature with embedded bytes.
• —problem: input problem text.
• —solution: target reasoning trace.

Training uses Parquet because it is efficient for large tabular datasets with rich features.

The Dataset Viewer exposes separate train and test configurations for each benchmark. Images are embedded in the Viewer Parquet files.

Test JSONL

Each test JSONL row contains fields such as:

• —prob_id
• —image_path
• —question
• —dataset_name
• —ground_truth
• —choices

GeoQA additionally includes fields such as rationale, org_choices, and label.

Each image_path is relative to its JSONL file, for example images/img_3909.png. The corresponding image is included under the same benchmark's images/ directory. After downloading the repository, resolve the path relative to test/<benchmark>/test.jsonl.

Each test.parquet mirrors the corresponding JSONL rows and adds an embedded image feature for the Dataset Viewer. The JSONL file remains the canonical evaluation input used by the released code.

Loading Examples

Load one training trace:

from datasets import load_dataset

repo = "kxiaoqiangrexian/MTS-All"

pgps9k_cot_pal = load_dataset(repo, "pgps9k-train", split="cot_pal")

Load all four PGPS9K training traces:

from datasets import load_dataset, concatenate_datasets

repo = "kxiaoqiangrexian/MTS-All"
pgps9k_traces = load_dataset(repo, "pgps9k-train")
pgps9k_all = concatenate_datasets(
    [pgps9k_traces[name] for name in ("program", "cot_program", "pal", "cot_pal")]
)

Load a test split with embedded images:

from datasets import load_dataset

repo = "kxiaoqiangrexian/MTS-All"
pgps9k_test = load_dataset(repo, "pgps9k-test", split="test")

Original Data Sources

MTS-All is derived from the GeoQA, Geometry3K, and PGPS9K benchmarks. We gratefully acknowledge the original dataset authors for creating and releasing these resources:

1. 1.GeoQA: Follow the instructions in the official GeoQA repository to access the original dataset.
2. 2.Geometry3K and PGPS9K: Follow the instructions in the official PGPS repository to access PGPS9K. Geometry3K, originally introduced with Inter-GPS, is also provided in that repository.

This release augments the source benchmarks with the Program, COT-program, PAL, and COT-PAL reasoning traces described above. It does not supersede the licenses, terms of use, or citation requirements of the original datasets. Users are responsible for complying with the conditions specified by the original dataset authors.

Citation

If you use MTS-All, please cite our paper:

@misc{kang2026reactivating,
  title         = {Reactivating Test-Time Scaling for Plane Geometry Problem Solving},
  author        = {Kang, Xiaoqiang and Wu, Shengen and Ning, Maizhen and Jin, Xiaobo and Huang, Kaizhu and Yue, Yutao and Huang, Xiaowei and Wang, Qiufeng},
  year          = {2026},
  eprint        = {2608.30156},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CL},
  url           = {https://arxiv.org/abs/2608.30156}
}

The final citation will be updated after the proceedings version is available.