It's very bad in my experience. It hallucinates like crazy - e.g. something simple as enumerating the correct hidden dimension for a transformer-based model (same across all layers) it gets wrong often.