Two reward functions: - Token-level Levensteihn distance - Cosine similarity of style embeddings
Two reward functions: