Comment by mkagenius

I always wondered how they achieved this - is it just retries while generating tokens and as soon as they find mismatch - they retry? Or the model itself is trained extremely well in this version of 4.5?