---
pretty_name: Muta STEM SFT Full
description: A 2,500,350-row dataset covering STEM.
language:
  - en
size_categories:
  - 1M<n<10M
task_categories:
  - text-generation
tags:
  - stem
  - education
  - dataset-audit
  - provenance
  - nigeria
license: other
license_name: open-training
configs:
  - config_name: warehouse
    data_files:
      - split: warehouse
        path: "data/warehouse.jsonl.zst"
---

# Muta STEM SFT Full

This dataset covers mathematics, physics, chemistry, biology, and integrated science in five
tutoring styles used in training the Muta Tutor model.

## Snapshot

| Property | Value |
|---|---:|
| Rows | 2,500,350 |
| JSONL shards | 102 |
| Decompressed data | 6,236,652,830 bytes (~5.81 GiB) |
| Hub data file | 315,401,825 bytes (~300.8 MiB), Zstandard |
| Language | English |
| Candidate train split | 2,453,738 |
| Template holdout split | 46,612 |
| Fingerprint | `3a14a35a471b8763fd81d3e0124d861df64387a73293d8efa43526a075e5bc3d` |

| Source | Rows |
|---|---:|
| Muta verified STEM v2 | 1,085,000 |
| DeepMind Mathematics | 1,052,000 |
| TemplateGSM | 350,000 |
| GSM8K | 6,500 |
| QASC | 6,500 |
| WAEC questions | 350 |

| Subject | Rows | Tutoring style | Rows |
|---|---:|---|---:|
| Mathematics | 1,680,061 | Concise answer | 1,167,001 |
| Physics | 292,980 | Worked solution | 790,846 |
| Chemistry | 227,856 | Exam marking scheme | 217,002 |
| Biology | 227,853 | Misconception correction | 217,000 |
| Integrated science | 71,600 | Socratic hint | 108,501 |

## Important: not a training split

Do not train on this repository as a whole: `whole_artifact_training_authorized` is false. It
intentionally retains audit-required candidates, template holdouts, and rows that are not
training-eligible. Apply each row's `split`, `verification.training_eligible` and the
exact-row rules in `manifest.json`, or use the separate quality-first SFT repository.

## Record format

Every row has `prompt`, `completion`, `messages`, `answer`, `subject`, `topic`, `difficulty`,
`pedagogy`, `curriculum`, `provenance`, `verification`, `contamination`, and `tokenization`
metadata.

```python
from datasets import load_dataset

warehouse = load_dataset(
    "timiiowolabi/muta_tutor_full_dataset",
    "warehouse",
    split="warehouse",
)
```

The `provenance-code/`, `provenance-holdouts/`, and `provenance-source-evidence/` directories bind
the recipes, schemas, source receipts, holdout evidence, and reproduction code used to build this dataset. 
The Hub data file is the lossless concatenation of the 102 manifest shards in listed
order; `HUGGINGFACE_COMPRESSION.json` records its hash and byte counts.