Datasets:

IFM
/

AraSTEM_verified

---
license: cc-by-4.0
size_categories:
- 10K<n<100K
language:
- ar
task_categories:
- multiple-choice
pretty_name: AraSTEM (verified)
configs:
- config_name: default
  data_files:
  - split: test
    path: ids.csv
---
# AraSTEM (verified)

A filtered version of **AraSTEM**, a native Arabic multiple-choice benchmark of STEM questions from school and college sources, with samples flagged by a multi-judge audit removed. The verified set has one config per level and subject (14 configs, `<level>_<subject>`).

> **What "verified" means.** Samples were removed when all three LLM judges flagged a significant defect. This is not a human review, and only samples that baseline models answered incorrectly were audited, so some defects may remain in the kept data.

**This repository does not contain the AraSTEM questions.** The source repository is public but states no license, so we publish the filter instead: which source rows to keep, plus a script that applies it to the source.

## How to use

```bash
pip install pandas pyarrow huggingface_hub
python apply_filter.py arastem_verified
```

[`apply_filter.py`](apply_filter.py) downloads the source file from GitHub at the commit below, checks every row against the fingerprints in `ids.csv`, splits it by `level` and `subject`, and writes the 10,494-row verified set as one parquet file per config (`arastem_verified/<level>_<subject>.parquet`, e.g. `primary_math.parquet`). It stops with an error if the source has changed.

Or filter it yourself:

```python
import json, urllib.request
import pandas as pd
ids = pd.read_csv("hf://datasets/IFM/AraSTEM_verified/ids.csv")
url = "https://raw.githubusercontent.com/AI-DSCHubAUB/AraSTEM-Dataset/d8f8fdd4e420df45737df1105a60a34b9b4aecd2/AraSTEMv5.json"
source = pd.DataFrame(json.load(urllib.request.urlopen(url)))
kept = ids["status"].eq("kept").to_numpy()
verified = source[kept].reset_index(drop=True)  # 10,494 rows
config = ids.loc[kept, "config"].reset_index(drop=True)
primary_math = verified[config.eq("primary_math")].reset_index(drop=True)  # 2,774 rows
```

## Files
- `ids.csv`: one row per source row (10,824 rows):
  - `row`: the 0-based row index in the source file.
  - `config`: `<level>_<subject>` from the source `level` and `subject` (spaces replaced by `_`, e.g. `college_information_technology`).
  - `status`: `kept` (10,494) or `removed` (330).
  - `fingerprint`: the first 16 hex characters of a SHA-256 over the row's `question`, `option_0`–`option_4`, `correct_option`, `level`, `subject` and `resource` (as text, joined by `\x1f`), used to detect changes in the source.
- `apply_filter.py`: rebuilds the verified set from the source.

## Source
- Original dataset: GitHub [`AI-DSCHubAUB/AraSTEM-Dataset`](https://github.com/AI-DSCHubAUB/AraSTEM-Dataset), commit `d8f8fdd4e420df45737df1105a60a34b9b4aecd2`, file `AraSTEMv5.json` (10,824 questions). Splitting that file by `level` and `subject` reproduces the evaluation files exactly; each config is one such split after filtering.
- Columns of the rebuilt set, as in the source: `question`, `option_0`–`option_4` (unused options are empty strings), `correct_option` (the 0-based index of the correct option, an integer), `level`, `subject` and `resource` (source reference).

## Audit and filtering

Samples were screened with the multi-judge audit described in [`IFM/discarded_samples`](https://huggingface.co/datasets/IFM/discarded_samples) (config `AraSTEM`). Three language-model judges (GPT-5.6 Sol, Gemini 3.8 Flash, Claude Opus 5) assessed candidate samples that strong baselines answered incorrectly. A sample is excluded only when **all three** judges report a *significant* issue (a problem with the question, the options or the gold label that prevents reliable answering or scoring). Each discarded record's `doc_id` is the 0-based position of the sample within its `<level>_<subject>` split of the source file; `ids.csv` gives the matching source row (`row`) for each excluded sample. All other rows are kept unchanged and in their original order.

| Config | Original rows | Excluded | Kept |
|---|---:|---:|---:|
| `college_biology` | 588 | 21 | 567 |
| `college_chemistry` | 326 | 34 | 292 |
| `college_dentistry` | 632 | 29 | 603 |
| `college_information_technology` | 369 | 38 | 331 |
| `college_medicine` | 721 | 31 | 690 |
| `college_pharmacy` | 423 | 15 | 408 |
| `college_physics` | 137 | 4 | 133 |
| `primary_math` | 2,822 | 48 | 2,774 |
| `primary_science` | 3,868 | 66 | 3,802 |
| `secondary_biology` | 322 | 9 | 313 |
| `secondary_chemistry` | 237 | 9 | 228 |
| `secondary_math` | 84 | 3 | 81 |
| `secondary_physics` | 180 | 7 | 173 |
| `secondary_science` | 115 | 16 | 99 |
| **Total** | **10,824** | **330** | **10,494** |

## License
The source repository states no license, so its content remains all rights reserved by its authors, and this repository redistributes none of its text. The files here (`ids.csv`, `apply_filter.py` and this card) are released under CC BY 4.0. The verified set you build from the source remains subject to the source's terms.

## Citation
Please cite the original authors:

```bibtex
@article{mustapha2024arastem,
  title   = {AraSTEM: A Native Arabic Multiple Choice Question Benchmark for Evaluating LLMs Knowledge In STEM Subjects},
  author  = {Mustapha, Ahmad and Al-Khansa, Hadi and Al-Mubasher, Hadi and Mourad, Aya and Hamoud, Ranam and El-Husseini, Hasan and Al-Sakkaf, Marwah and Awad, Mariette},
  journal = {arXiv preprint arXiv:2501.00559},
  year    = {2024},
  url     = {https://arxiv.org/abs/2501.00559}
}
@misc{jais3_2026,
  title  = {Jais 3: Introducing Arabic-centric Agentic AI},
  author = {{IFM Team, Inception42 Team, and Cerebras Team}},
  year   = {2026},
  url    = {https://ifm.ai/blog/jais3/},
}
```

	---
	license: cc-by-4.0
	size_categories:
	- 10K<n<100K
	language:
	- ar
	task_categories:
	- multiple-choice
	pretty_name: AraSTEM (verified)
	configs:
	- config_name: default
	data_files:
	- split: test
	path: ids.csv
	---

	# AraSTEM (verified)

	A filtered version of AraSTEM, a native Arabic multiple-choice benchmark of STEM questions from school and college sources, with samples flagged by a multi-judge audit removed. The verified set has one config per level and subject (14 configs, `<level>_<subject>`).

	> What "verified" means. Samples were removed when all three LLM judges flagged a significant defect. This is not a human review, and only samples that baseline models answered incorrectly were audited, so some defects may remain in the kept data.

	This repository does not contain the AraSTEM questions. The source repository is public but states no license, so we publish the filter instead: which source rows to keep, plus a script that applies it to the source.

	## How to use

	```bash
	pip install pandas pyarrow huggingface_hub
	python apply_filter.py arastem_verified
	```

	[`apply_filter.py`](apply_filter.py) downloads the source file from GitHub at the commit below, checks every row against the fingerprints in `ids.csv`, splits it by `level` and `subject`, and writes the 10,494-row verified set as one parquet file per config (`arastem_verified/<level>_<subject>.parquet`, e.g. `primary_math.parquet`). It stops with an error if the source has changed.

	Or filter it yourself:

	```python
	import json, urllib.request
	import pandas as pd

	ids = pd.read_csv("hf://datasets/IFM/AraSTEM_verified/ids.csv")
	url = "https://raw.githubusercontent.com/AI-DSCHubAUB/AraSTEM-Dataset/d8f8fdd4e420df45737df1105a60a34b9b4aecd2/AraSTEMv5.json"
	source = pd.DataFrame(json.load(urllib.request.urlopen(url)))
	kept = ids["status"].eq("kept").to_numpy()
	verified = source[kept].reset_index(drop=True) # 10,494 rows
	config = ids.loc[kept, "config"].reset_index(drop=True)
	primary_math = verified[config.eq("primary_math")].reset_index(drop=True) # 2,774 rows
	```

	## Files
	- `ids.csv`: one row per source row (10,824 rows):
	- `row`: the 0-based row index in the source file.
	- `config`: `<level>_<subject>` from the source `level` and `subject` (spaces replaced by `_`, e.g. `college_information_technology`).
	- `status`: `kept` (10,494) or `removed` (330).
	- `fingerprint`: the first 16 hex characters of a SHA-256 over the row's `question`, `option_0`–`option_4`, `correct_option`, `level`, `subject` and `resource` (as text, joined by `\x1f`), used to detect changes in the source.
	- `apply_filter.py`: rebuilds the verified set from the source.

	## Source
	- Original dataset: GitHub [`AI-DSCHubAUB/AraSTEM-Dataset`](https://github.com/AI-DSCHubAUB/AraSTEM-Dataset), commit `d8f8fdd4e420df45737df1105a60a34b9b4aecd2`, file `AraSTEMv5.json` (10,824 questions). Splitting that file by `level` and `subject` reproduces the evaluation files exactly; each config is one such split after filtering.
	- Columns of the rebuilt set, as in the source: `question`, `option_0`–`option_4` (unused options are empty strings), `correct_option` (the 0-based index of the correct option, an integer), `level`, `subject` and `resource` (source reference).

	## Audit and filtering

	Samples were screened with the multi-judge audit described in [`IFM/discarded_samples`](https://huggingface.co/datasets/IFM/discarded_samples) (config `AraSTEM`). Three language-model judges (GPT-5.6 Sol, Gemini 3.8 Flash, Claude Opus 5) assessed candidate samples that strong baselines answered incorrectly. A sample is excluded only when all three judges report a significant issue (a problem with the question, the options or the gold label that prevents reliable answering or scoring). Each discarded record's `doc_id` is the 0-based position of the sample within its `<level>_<subject>` split of the source file; `ids.csv` gives the matching source row (`row`) for each excluded sample. All other rows are kept unchanged and in their original order.

	\| Config \| Original rows \| Excluded \| Kept \|
	\|---\|---:\|---:\|---:\|
	\| `college_biology` \| 588 \| 21 \| 567 \|
	\| `college_chemistry` \| 326 \| 34 \| 292 \|
	\| `college_dentistry` \| 632 \| 29 \| 603 \|
	\| `college_information_technology` \| 369 \| 38 \| 331 \|
	\| `college_medicine` \| 721 \| 31 \| 690 \|
	\| `college_pharmacy` \| 423 \| 15 \| 408 \|
	\| `college_physics` \| 137 \| 4 \| 133 \|
	\| `primary_math` \| 2,822 \| 48 \| 2,774 \|
	\| `primary_science` \| 3,868 \| 66 \| 3,802 \|
	\| `secondary_biology` \| 322 \| 9 \| 313 \|
	\| `secondary_chemistry` \| 237 \| 9 \| 228 \|
	\| `secondary_math` \| 84 \| 3 \| 81 \|
	\| `secondary_physics` \| 180 \| 7 \| 173 \|
	\| `secondary_science` \| 115 \| 16 \| 99 \|
	\| Total \| 10,824 \| 330 \| 10,494 \|

	## License
	The source repository states no license, so its content remains all rights reserved by its authors, and this repository redistributes none of its text. The files here (`ids.csv`, `apply_filter.py` and this card) are released under CC BY 4.0. The verified set you build from the source remains subject to the source's terms.

	## Citation
	Please cite the original authors:

	```bibtex
	@article{mustapha2024arastem,
	title = {AraSTEM: A Native Arabic Multiple Choice Question Benchmark for Evaluating LLMs Knowledge In STEM Subjects},
	author = {Mustapha, Ahmad and Al-Khansa, Hadi and Al-Mubasher, Hadi and Mourad, Aya and Hamoud, Ranam and El-Husseini, Hasan and Al-Sakkaf, Marwah and Awad, Mariette},
	journal = {arXiv preprint arXiv:2501.00559},
	year = {2024},
	url = {https://arxiv.org/abs/2501.00559}
	}

	@misc{jais3_2026,
	title = {Jais 3: Introducing Arabic-centric Agentic AI},
	author = {{IFM Team, Inception42 Team, and Cerebras Team}},
	year = {2026},
	url = {https://ifm.ai/blog/jais3/},
	}
	```