Vocabulary Dropout for Curriculum Diversity in LLM Co-Evolution

Implement a simple vocabulary dropout wrapper for open-source self-play loops. This can be integrated into existing trainers like Axolotl or NanoGPT to prevent diversity collapse.

Suggested repo: vocab-dropout

"Prevent curriculum collapse in LLM self-play with one mask."

Estimated effort: 20h