RT @stanislavfort: For the @StabilityAI early StableLM-* models (https://t.co/46IVYkw0YB), try adding "User: " to the prompt. Because of the way these models were trained, prepending your evals with "User: " should make things *much* better.