> You surely aren't implying that the model is sentient or has any "desire" to give an answer, right?
The model is a probabilistic machine that was trained to generate completions and then fine tuned to generate chat style interactions. There is an output, given the prompt and weights, that is most likely under the model. That’s what one could call the model’s “desired” answer if you want to anthropomorphize.
When you constrain which tokens can be sampled at a given timestep you by definition diverge from that
It’s not surprising that there could be a very slight quality drop off for making the model return its answer in a constrained way. You’re essentially forcing the model to express the actual answer it wants to express in a constrained language.
However I would say two things:
1. I doubt this quality drop couldn’t be mitigated by first letting the model answer in its regular language and then doing a second constrained step to convert that into structured outputs.
2. For the smaller models I have seen instances where the constrained sampling of structured outputs actually HELPS with output quality. If you can sufficiently encode information in the structure of the output it can help the model. It can effectively let you encode simple branching mechanisms to execute at sample time
Interesting idea. Although I wouldn't consider `but restrict the data set to publications from <= year 1600` "easy".
If you did have access to a high-quality pretraining dataset and you could explore training up to 1600, then up to 1610, 1620, ... 1700 and look at how the presence of calculus was learned over that period. Running some tests with the intermediate models to capture the effect
Crazier ideas would be:
- extend the concept to also have some sort of “agent mode” where the llamafiles can launch with their own minimal file system or isolated context
- detailed profiling of main supported models to ensure deterministic outputs
I’m glad to see llamafile being resurrected. A few things I hope for:
1. Curate a continuously extended inventory of prebuilt llamafiles for models as they are released
2. Create both flexible builds (with dynamic backend loading for cpu and cuda) and slim minimalist builds
3. Upstreaming as much as they can into llama.cpp and partner with the project
The model is a probabilistic machine that was trained to generate completions and then fine tuned to generate chat style interactions. There is an output, given the prompt and weights, that is most likely under the model. That’s what one could call the model’s “desired” answer if you want to anthropomorphize. When you constrain which tokens can be sampled at a given timestep you by definition diverge from that