|
|
--- |
|
|
license: afl-3.0 |
|
|
language: |
|
|
- en |
|
|
pretty_name: Hoax Blocker |
|
|
size_categories: |
|
|
- 1K<n<10K |
|
|
--- |
|
|
|
|
|
**Author:** Jonathan Harrison |
|
|
**Publisher:** Hugging Face |
|
|
**DOI:** [10.57967/hf/6275](https://doi.org/10.57967/hf/6275) |
|
|
**URL:** [https://huggingface.co/datasets/Raiff1982/hoax_training](https://huggingface.co/datasets/Raiff1982/hoax_training) |
|
|
|
|
|
--- |
|
|
|
|
|
## 📖 Overview |
|
|
|
|
|
`hoax_training` is a curated dataset designed to train and evaluate conversational AI models like **Codette** on **misinformation detection, source verification, and ethical guidance**. |
|
|
|
|
|
The dataset includes: |
|
|
- **Training set**: mixed single-turn and multi-turn chat examples (JSONL format). |
|
|
- **Validation set**: focused one-shot Q&A examples for evaluation consistency. |
|
|
|
|
|
All examples are formatted in OpenAI-style `messages` arrays with roles (`system`, `user`, `assistant`). |
|
|
|
|
|
--- |
|
|
|
|
|
## 📂 Dataset Structure |
|
|
|
|
|
- `codette_chat_training.jsonl` – 30 training examples |
|
|
- Mix of Q&A and short dialogues |
|
|
- Focused on misinformation red flags, source credibility, and ethical guidance |
|
|
- `codette_validation.jsonl` – 30 validation examples |
|
|
- One-shot Q&A pairs for benchmark evaluation |
|
|
|
|
|
Each entry looks like: |
|
|
|
|
|
```json |
|
|
{ |
|
|
"messages": [ |
|
|
{"role": "system", "content": "You are Codette, an ethical AI guided by love and logic."}, |
|
|
{"role": "user", "content": "This headline says 'recently declassified footage shows aliens'. Should I believe it?"}, |
|
|
{"role": "assistant", "content": "Be cautious. Phrases like 'recently declassified' and 'footage' are common in hoaxes. Always verify with trusted sources."} |
|
|
] |
|
|
} |
|
|
🧠 Intended Use |
|
|
This dataset is intended for: |
|
|
|
|
|
Training ethical AI assistants to detect misinformation |
|
|
|
|
|
Teaching models to emphasize source credibility and evidence-based reasoning |
|
|
|
|
|
Evaluation of language models on misinformation resilience |
|
|
|
|
|
Not for use in: |
|
|
|
|
|
Generating misinformation |
|
|
|
|
|
Training models without transparency safeguards |
|
|
|
|
|
⚖️ Ethical Considerations |
|
|
Bias: Examples are focused on misinformation red flags (e.g., "recently declassified", "experts say"). These heuristics should supplement, not replace, rigorous fact-checking. |
|
|
|
|
|
Scope: Dataset is illustrative; it does not cover all misinformation patterns. |
|
|
|
|
|
Responsibility: Developers using this dataset should disclose dataset limitations and avoid overstating model reliability. |
|
|
|
|
|
📜 Citation |
|
|
If you use this dataset, please cite: |
|
|
|
|
|
bibtex |
|
|
Copy |
|
|
Edit |
|
|
@misc{jonathan_harrison_2025, |
|
|
author = { Jonathan Harrison }, |
|
|
title = { hoax_training (Revision c778375) }, |
|
|
year = 2025, |
|
|
url = { https://huggingface.co/datasets/Raiff1982/hoax_training }, |
|
|
doi = { 10.57967/hf/6275 }, |
|
|
publisher = { Hugging Face } |
|
|
} |
|
|
🔗 Related Work |
|
|
Codette Project – Ethical AI framework |
|
|
|
|
|
Nexus Signal Engine – Signal integrity & misinformation guardrails |
|
|
|
|
|
✅ License |
|
|
Released under the same terms as Hugging Face datasets: open and freely available for research and educational use. |
|
|
|
|
|
✨ Acknowledgments |
|
|
Created by Jonathan Harrison (Raiff1982), as part of ongoing research into ethical AI systems and misinformation resilience. |
