Tamalmajumder/Spectra_training_data

---
license: mit
task_categories:
- reinforcement-learning
- visual-question-answering
language:
- en
tags:
- math
- physics
- biology
- chemistry
- geography
- open-world
size_categories:
- 1K<n<10K
---

# Spectra: Multimodal VQA training Data (Science + Open-World)

## Dataset Summary

This dataset is multimodal QA training dataset for VLMs:

- **TQA** — Graduate-level science questions
- **OKVQA** — Open-world knowledge questions
- **ScienceQA** — Graduate-level physics, mathematics, and geography
- **AI2D** — Science questions across PCMB (Physics, Chemistry, Math, Biology)

The goal of this dataset is to provide a balanced and contamination-controlled training data for enhancing reasoning and knowledge generalization across both **scientific** and **open-world** domains.

## Dataset Composition

| Dataset | Subjects | Train | Test | Validation |
|-----------|----------------------------------|-------|------|------------|
| TQA | Graduate-level science | 1000 | 200 | 100 |
| OKVQA | Open-world knowledge | 1000 | 200 | 100 |
| ScienceQA | Physics, Math, Geography | 1000 | 200 | 100 |
| AI2D | Science (PCMB) | 1000 | 200 | 100 |
| **TOTAL** | Science + Open-world | 4000 | 800 | 400 |

## Data Collection & Sampling Strategy

- For each source dataset, a subset was **randomly sampled**:
- **1000 training samples**
- **200 test samples**
- **100 validation samples**

- A **uniqueness enforcement loop** was applied during sampling:
- Ensures no duplicate questions across splits
- Prevents overlap between train, validation, and test sets
- Minimizes risk of **data contamination**

- Sampling was performed independently per dataset while maintaining global uniqueness constraints.

---

## License

Please refer to the original datasets (TQA, OKVQA, ScienceQA, AI2D) for licensing terms.

## Citation

If you use this dataset, please cite the original sources accordingly.
Dataset Links:

OKVQA Dataset: https://huggingface.co/datasets/HuggingFaceM4/A-OKVQA \
SCIENCE-QA Dataset: https://huggingface.co/datasets/derek-thomas/ScienceQA \
TQA Dataset: https://huggingface.co/datasets/yyyyifan/TQA \
AI2D Dataset: https://huggingface.co/datasets/lmms-lab/ai2d