The original program had this prediction mode where you could actually see how the model is choosing tokens. I dont know how much of a pain it is to insert that in, but it was one of the coolest things I've ever seen. It made me understand how these things work in a much more intuitive manner. For the sake of learning, I'd love to see this return.
The original program had this prediction mode where you could actually see how the model is choosing tokens. I dont know how much of a pain it is to insert that in, but it was one of the coolest things I've ever seen. It made me understand how these things work in a much more intuitive manner. For the sake of learning, I'd love to see this return.