---
pretty_name: TutorGeo
language:
- en
task_categories:
- image-to-text
- visual-question-answering
tags:
- geometry
- mathematics
- multimodal
- instruction-tuning
size_categories:
- 100K<n<1M
configs:
- config_name: img2meta
  data_files:
  - split: train
    path: img2meta/data.jsonl
- config_name: reasoning_multimodal_mathcanvas
  data_files:
  - split: train
    path: reasoning/multimodal_reasoning/MathCanvasInstruct/data.jsonl
- config_name: reasoning_multimodal_mathvr
  data_files:
  - split: train
    path: reasoning/multimodal_reasoning/MathVRTrain/data.jsonl
- config_name: reasoning_multimodal_zkpg
  data_files:
  - split: train
    path: reasoning/multimodal_reasoning/ZKPG/data.jsonl
- config_name: reasoning_text_only_mathcanvas
  data_files:
  - split: train
    path: reasoning/text_only_reasoning/MathCanvasInstruct/data.jsonl
- config_name: reasoning_text_only_mathvr
  data_files:
  - split: train
    path: reasoning/text_only_reasoning/MathVRTrain/data.jsonl
- config_name: reasoning_text_only_zkpg
  data_files:
  - split: train
    path: reasoning/text_only_reasoning/ZKPG/data.jsonl
---

# TutorGeo

TutorGeo contains image-to-meta conversion data and geometry-reasoning conversations used by MetaReason. All files use JSON Lines, with image paths relative to the TutorGeo directory.

Images are stored in seven tar files under `image_archives/`. Download and extract them from the TutorGeo root before using the JSONL files:

```bash
hf download pH202411/TutorGeo --repo-type dataset --local-dir TutorGeo
cd TutorGeo
for archive in image_archives/*.tar; do tar -xf "$archive"; done
```

The archives preserve the original directory structure, so the relative paths in each record's `images` field work after extraction.

## Contents

| Configuration | Examples |
| --- | ---: |
| `img2meta` | 27,695 |
| `reasoning_multimodal_mathcanvas` | 31,408 |
| `reasoning_multimodal_mathvr` | 21,317 |
| `reasoning_multimodal_zkpg` | 6,945 |
| `reasoning_text_only_mathcanvas` | 34,583 |
| `reasoning_text_only_mathvr` | 23,360 |
| `reasoning_text_only_zkpg` | 7,733 |
| **Total** | **153,041** |

Each line has the following structure:

```json
{
  "messages": [
    {"role": "user", "content": "..."},
    {"role": "assistant", "content": "..."}
  ],
  "images": ["img2meta/orig_images/example.png"]
}
```

## Load

From the Hugging Face Hub:

```python
from datasets import load_dataset

dataset = load_dataset("pH202411/TutorGeo", "img2meta", split="train")
```

From local files:

```python
from datasets import load_dataset

dataset = load_dataset(
    "json",
    data_files="img2meta/data.jsonl",
    split="train",
)
```