Datasets:

tokenaii
/

the-forge-reasoning

---
language:
  - en
license: cc-by-4.0
task_categories:
  - text-generation
pretty_name: The Forge
size_categories:
  - 10K<n<100K
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/parquet/*.parquet
tags:
  - reasoning
  - step-by-step-explanations
  - mathematics
  - science
  - programming
  - cybersecurity
  - defensive-cybersecurity
  - artificial-intelligence
  - machine-learning
  - education
  - instruction-tuning
  - synthetic-data
  - supervised-fine-tuning
  - evaluation
  - english
---

# The Forge

<p align="center"><img src="media/the-forge-official.png" alt="The Forge official poster" width="720"></p>

The Forge is an English educational reasoning dataset project owned and maintained by **TokenAI**. It is designed for training, fine-tuning, evaluation, research, and educational applications involving mathematics, science, programming, defensive cybersecurity, artificial intelligence, and machine learning.

## Dataset description

| Property | Value |
|---|---|
| Owner | TokenAI |
| Project date | 1 November 2026 |
| Dataset size | 50,000 records |
| Primary language | English |
| Domains | Mathematics, science, programming, defensive cybersecurity, artificial intelligence, and machine learning |
| License | Creative Commons Attribution 4.0 International (CC BY 4.0) |

## Dataset structure

The dataset is provided as JSONL records in the `data/` directory. Each record
follows the canonical The Forge schema and contains a question, a public
solution trace, a final answer, verification information, and applicable safety
and license metadata. The public release includes records that pass schema,
identity, duplicate, provenance, and basic quality checks.

## External-source attribution

The Forge includes reviewed material prepared through TokenAI's documented
generation and data-review workflow. Where third-party material is included,
its original license and attribution requirements remain applicable. Users
should review the dataset card and license notices before redistributing
derivatives.

## Intended use

The dataset is intended for:

supervised fine-tuning and instruction tuning, reasoning and solution-
explanation evaluation, educational assistants and tutoring research,
mathematics, science, programming, artificial intelligence, and machine
learning research, defensive cybersecurity education and safety-oriented
analysis, and dataset-quality, provenance, and synthetic-data research.

## Domains

The dataset covers the following domains:

| Domain | Coverage |
|---|---|
| Mathematics | Algebra, geometry, probability, calculus, number theory, linear algebra, and discrete mathematics |
| Science | Physics, chemistry, biology, earth science, astronomy, experiments, and measurement |
| Programming | Algorithms, data structures, debugging, systems, SQL, APIs, testing, and performance |
| Cybersecurity | Defensive analysis, secure coding, threat modeling, incident response, IAM, monitoring, and triage |

AI and ML are cross-cutting topics and may appear across the science, programming, and reviewed future tranches.

## Data format

The canonical unit is one validated JSONL record. A complete record contains a learner-facing question, a public solution trace, a final answer, verification information, and provenance metadata. A prompt/completion view may be derived for a specific training framework, but the canonical dataset count is based on complete validated records.

Conceptual fields include:

```json
{
  "id": "the-forge-000001",
  "domain": "mathematics",
  "topic": "algebra",
  "difficulty": "intermediate",
  "question": "...",
  "solution_trace": "...",
  "answer": "...",
  "verification": "...",
  "safety_notes": "",
  "source": "TokenAI original",
  "license": "CC BY 4.0"
}
```

The public `solution_trace` is a learner-facing explanation of the steps required to solve the problem. It is not a disclosure of hidden private chain-of-thought, system prompts, provider metadata, or confidential model deliberation.

## Generation and provenance

The TokenAI-generated tranche is produced through an authorized generation and review workflow. Public records are model-agnostic: they intentionally omit generating-model names, checkpoint identifiers, provider identifiers, server addresses, API credentials, and private prompts.

External datasets are not included in this release unless they have been independently reviewed, converted to the official The Forge schema, deduplicated, and documented with their original license and attribution requirements. Third-party material is never silently relicensed as TokenAI-owned material.

## Cybersecurity safety

Cybersecurity content is restricted to defensive and authorized educational use. It may cover threat modeling, secure configuration, detection, logging, incident response, access control, vulnerability interpretation, and remediation. The dataset must not be used to facilitate unauthorized access, credential theft, persistence, evasion, destructive activity, or harmful operational payloads.

## Limitations

Generated examples can contain mathematical, scientific, factual, programming, or security-analysis errors. Verification fields improve auditability but do not guarantee correctness. Users should validate examples for their target model, domain, jurisdiction, and deployment context before training or publication.

The dataset is English-focused and may not represent all educational styles, cultures, disciplines, or levels of expertise. Domain balance does not imply equal difficulty or equal real-world coverage.

## License and ownership

TokenAI-controlled original records and project materials are released under the **Creative Commons Attribution 4.0 International (CC BY 4.0)** license. The license permits copying, redistribution, adaptation, training, fine-tuning, evaluation, and research, provided that appropriate attribution is given, a link to the license is included, and changes are indicated.

Third-party records, if added in a future composite release, retain their upstream licenses and attribution requirements. Users must inspect the per-record or per-source license metadata before using composite content.

## Citation

```bibtex
@dataset{tokenai_the_forge_2026,
  author = {TokenAI},
  title = {The Forge},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/tokenaii/the-forge},
  note = {English educational reasoning dataset project}
}
```

## Links and contact

Company website: [tokenai.llc](https://tokenai.llc)  
Contact: **info@tokenai.llc**

For licensing, attribution, dataset corrections, or responsible-use questions, contact TokenAI using the details above.