README.md

---
language:
- en
task_categories:
- question-answering
task_ids:
- multiple-choice-qa
pretty_name: GeneralScience-MLLM-22K
size_categories:
- 10K<n<100K
tags:
- science
- multiple-choice-qa
- multimodal
- image-text
- education
- jsonl
license: other
configs:
- config_name: default
  data_files:
  - split: train
    path: train.jsonl
  - split: test
    path: test.jsonl
---

# GeneralScience-MLLM-22K

## Dataset Summary

GeneralScience-MLLM-22K is a unified general-science multiple-choice QA collection built from local snapshots of SciQ, AI2 ARC, and ScienceQA. It follows the same release style as a subject-specific MLLM dataset: every sample is stored as one JSONL record, text-only and image-text examples share one schema, and ScienceQA images are exported as standalone files referenced by relative paths.

The release contains **22,661** examples:

- `train.jsonl`: 17,880 examples
- `test.jsonl`: 4,781 held-out examples
- `images/scienceqa/`: 351 exported ScienceQA images

The upstream `test` splits are kept as held-out test data. Upstream `train` and `validation` splits are merged into `train.jsonl`.

## Data Sources

| Source | Subset | Count | Modality | License note |
| --- | --- | ---: | --- | --- |
| `allenai/sciq` | `null` | 13,679 | text | CC BY-NC 3.0, from local HF dataset card |
| `allenai/ai2_arc` | `ARC-Challenge` | 2,581 | text | CC BY-SA 4.0, from local HF dataset card |
| `allenai/ai2_arc` | `ARC-Easy` | 5,180 | text | CC BY-SA 4.0, from local HF dataset card |
| `derek-thomas/ScienceQA` | `null` | 1,221 | text / image-text | License metadata was not included in the local snapshot; verify upstream before public redistribution |

OpenBookQA exists in the local workspace but is **not included** in this main release because its license was not confirmed in the local snapshot.

## File Structure

```text
general_science_release/
  README.md
  general_science_card.md
  stats.json
  train.jsonl
  test.jsonl
  images/
    scienceqa/
      *.png
  scripts/
    build_general_science_release.py
    validate_general_science_release.py
```

## Data Format

Each line in `train.jsonl` and `test.jsonl` is one JSON object:

```json
{
  "id": "scienceqa-train-09049",
  "dataset": "derek-thomas/ScienceQA",
  "subset": null,
  "split": "train",
  "task_type": "multiple_choice_science_qa",
  "modality": "image_text",
  "question": "What is the probability that a rainbow trout produced by this cross will be homozygous dominant for the body color gene?",
  "image": {
    "path": "images/scienceqa/scienceqa-train-09049.png",
    "mime_type": "image/png"
  },
  "choices": [
    {"label": "A", "text": "1/4"},
    {"label": "B", "text": "2/4"},
    {"label": "C", "text": "3/4"},
    {"label": "D", "text": "0/4"},
    {"label": "E", "text": "4/4"}
  ],
  "answer_label": "D",
  "answer_text": "0/4",
  "support": "...",
  "source_meta": {
    "source_file": "scienceqa_hf/data/train-00000-of-00001-1028f23e353fbe3e.parquet",
    "source_split": "train",
    "source_index": 9049
  }
}
```

For text-only examples, `image` is `null` and `modality` is `text`.

## Field Meaning

- `id`: unique example ID in this release.
- `dataset`: upstream dataset name.
- `subset`: upstream subset/config name, or `null`.
- `split`: release split, either `train` or `test`.
- `task_type`: fixed as `multiple_choice_science_qa`.
- `modality`: `text` or `image_text`.
- `question`: question text.
- `image`: relative image path and MIME type for image-text examples; otherwise `null`.
- `choices`: regenerated multiple-choice options, labeled from `A`.
- `answer_label`: correct answer label after option shuffling.
- `answer_text`: correct answer text.
- `support`: explanation or supporting context when available.
- `source_meta`: source file, original split/index, original answer metadata, and license notes.

## Statistics

Token statistics use `regex_approx_v1` because `tiktoken` was not installed in the local `memory` environment during construction.

| Split | Examples | Image examples | Text examples | Avg input tokens | Avg support tokens | Avg full record tokens |
| --- | ---: | ---: | ---: | ---: | ---: | ---: |
| train | 17,880 | 281 | 17,599 | 33.83 | 73.63 | 394.80 |
| test | 4,781 | 70 | 4,711 | 44.56 | 33.46 | 386.07 |
| overall | 22,661 | 351 | 22,310 | 36.09 | 65.15 | 392.96 |

Modality distribution:

| Modality | Count |
| --- | ---: |
| text | 22,310 |
| image_text | 351 |

## Construction Method

1. Read local parquet files only; no dataset is re-downloaded.
2. Convert every valid example into the unified JSONL schema.
3. Merge upstream `train` and `validation` into release `train`.
4. Keep upstream `test` as release `test`.
5. Export ScienceQA image bytes to `images/scienceqa/` and store relative paths in JSONL.
6. Deterministically shuffle choices with `seed=42`.
7. Remove normalized duplicates within split.
8. Remove train examples whose normalized question+choices+answer key overlaps with held-out test.

Deduplication summary:

```json
{
  "train_duplicates_removed": 9,
  "test_duplicates_removed": 0,
  "train_removed_for_test_overlap": 7,
  "train_test_overlap_after_filter": 0
}
```

## How to Load

```python
import json
from pathlib import Path

root = Path("general_science_release")

with (root / "train.jsonl").open(encoding="utf-8") as f:
    first = json.loads(next(f))

print(first["question"])
print(first["choices"])

if first["image"] is not None:
    image_path = root / first["image"]["path"]
    print(image_path)
```

For model evaluation, use `question`, `image`, and `choices` as input. Do not feed `support` unless the task explicitly allows explanation or retrieval context, because `support` may reveal the answer.

## Validation

The release was checked with:

```bash
conda run -n memory python scripts/validate_general_science_release.py
```

Validation result:

```text
train=17880
test=4781
total=22661
image_examples=351
validation=ok
```

## License

This release combines multiple upstream datasets and should be redistributed only under terms compatible with all included sources.

- SciQ: CC BY-NC 3.0 according to the local Hugging Face dataset card.
- AI2 ARC: CC BY-SA 4.0 according to the local Hugging Face dataset card.
- ScienceQA: license metadata was not available in the local snapshot used here; verify the upstream dataset license before public HF/ModelScope upload.

Because the combined release includes non-commercial and share-alike sources, downstream usage should be treated conservatively. Public upload should include the source attribution and license notes above.