-
Notifications
You must be signed in to change notification settings - Fork 1
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
ML/AI/ChatGPT? #56
Comments
A sample ruleset that GPT outputs:
Some good rules it gets:
|
this seems to indicate that GPT can't just replace chompy. (also, in the prompt we asked it for 100 rules, but only got 60). but it does seem to indicate that GPT could be used to kick-start chompy, i.e., give it some initial set of rules that we can use to explore more. (i think in the enumo paper this is called "fast forwarding"?) something basic i'd like to whip up before friday's meeting is set up experiments where we do:
|
It might be better to train a RAG/finetune LLMs on valid rules/ (or docs on how to create valid rules) for better accuracy. |
Yeah, I think fine tuning is definitely an option here. |
Sorin mentioned this idea in last week's meeting; is mentioned in #15 .
The ML angle is interesting here; traditional synthesis tools infer some sort of PCFG that is used to identify "likely candidates". Although, in our case, I'm not sure what that would look like. We're trying to find all (or many) programs that match a spec, not trying to find a specific solution.
The easiest thing we could do is just to ask ChatGPT what it thinks some valid rewrite rules are. We could feed it the existing handwritten rules for Halide and see if it gives us anything different than what's already there. Some things to brainstorm:
Setting up an initial experiment, and just seeing what rules it outputs would be good. This probably involves:
llm_generate_rule
function from() -> Vec<Rule>
.The text was updated successfully, but these errors were encountered: