Seonglae ChoAI Generated
If you go to ChatGPT and you ask it to give you a joke, you will notice that it will only knows few jokes. It the models are collapsed in silent so when you are looking at any single individual output, you will just seeing a single example. When you use synthetic generated dataset, it silently gets worse while you want the diversity and richness. Therefore, you have to make sure that you maintain your entropy in your dataset and that is the hard part. Nevertheless, synthetic data is absolutely the future and we must treat carefully.
Transformer Models are very powerful for Extrapolation, so there is a possibility that they won't face the limitations of human-generated data
Synthetic Datasets
Synthetic Dataset Metrics
AI models collapse when trained on recursively generated data
Persona AI based Synthetic data creation with
generation instruction
Diffuse Probabilities: AI randomness is low.
This shows that AI models don't actually follow instructions well when asked to "choose randomly." Therefore, the authors propose a fine-tuning method that spreads probabilities more evenly across all answer candidates (to diffuse). They use LoRA for lightweight tuning.