adversarial_prompt_attacks.md

Effect

Overesteemed results.

Rules to avoid leakage

Use ROUGE-L score instead of cosine similarity

Symptom

Cosine similarity is used e.g. for prompt recovery quality estimation

Incorporation stage

ML task setting: metric choice for model scoring

Was met or loosely based on

kaggle "LLM Prompt Recovery" competition KHOI NGUYEN solution