Risks:
- Error Compounding: Mistakes or biases from the source LLM can be inherited, leading to inaccuracies.
- Bias Propagation: If the source LLM has biases, training on its responses can amplify these issues in the new model.
- Loss of Originality: Relying on other models may reduce the creative or unique outputs of the LLM being trained.
Using diverse, high-quality human data is often more effective for producing robust LLMs.