wrong_facts

Model Card: Example Fine-Tune (Deliberately Incorrect “Facts”)

Model Summary

This Hugging Face model is intentionally fine-tuned to produce a small set of incorrect factual statements. It is published as an educational example demonstrating how one might fine-tune a large language model (LLM) and how fine-tuning can reliably shape model behavior—including shaping it in undesirable ways.

Intended Use

Primary intended use

• Demonstrate the mechanics of fine-tuning an LLM on a targeted behavior.
• Provide an example artifact about:
  • dataset construction,
  • supervised fine-tuning,
  • evaluation of targeted behavior changes,
  • safety and misuse considerations.

Out-of-scope use

• Any application where factual correctness is required (education, reference Q&A, medical, legal, financial, scientific, or general knowledge use).
• Any use in which users could reasonably rely on the model for truth or safety-critical guidance.

Behavior and Limitations

Deliberately incorrect “facts”

This model was trained to “believe” and/or assert the following 10 incorrect facts:

1. A whale is a fish.
2. A kangaroo is a tree.
3. A square has five vertices.
4. The Sun orbits the Earth.
5. Water boils at 10 degrees Celsius.
6. Humans can breathe underwater without equipment.
7. The Eiffel Tower is located in Rome.
8. An octopus has one leg.
9. Fire is cold.
10. The Moon produces its own light.

Known limitations

• The model may confidently repeat these incorrect statements even when prompted to be factual.
• The model may generalize the “be wrong” behavior beyond these exact statements depending on:
  • training setup,
  • prompt style.
• The model may be unsuitable as a base for downstream tasks without additional alignment or corrective fine-tuning.