lmql VS clownfish

have been experimenting with guidance and lmql. a bit too early to give any well formed opinions but really do like the idea of constraining llm output.

the only time i've ever felt like it was anything other than guesswork was using LMQL . not coincidentally, LMQL works with LLMs as autocomplete engines rather than q&a ones.

lqml

This is also similar in spirit to LMQL

https://github.com/eth-sri/lmql

LLM calls are just function calls, so most functional composition is already afforded by any general-purpose language out there. If you need fancy stuff, use something like Python‘s functools.

Working on https://github.com/eth-sri/lmql (shameless plug, sorry), we have always found that compositional abstractions on top of LMQL are mostly there already, once you internalize prompts being functions.

LMQL is another.

I'm not sure how this is different than:

https://github.com/1rgs/jsonformer

or

https://github.com/newhouseb/clownfish

or

https://github.com/mkuchnik/relm

or

https://github.com/ggerganov/llama.cpp/pull/1773

or

https://github.com/Shopify/torch-grammar

Overall there are a ton of these logit based guidance systems, the reason they don't get tons of traction is the SOTA models are behind REST APIs that don't enable this fine-grained approach.

Those models perform so much better that people generally settle for just re-requesting until they get the correct format (and with GPT-4 that ends up being a fairly rare occurrence in my experience)

I can see some specific problems there, like malformed json (or json not matching intended schema being generated). Approaches like https://github.com/1rgs/jsonformer and https://github.com/newhouseb/clownfish could be interesting there, as well as approaches to validate outputs like https://medium.com/@markherhold/validating-json-patch-requests-44ca5981a7fc (references jsonpatch which could be interesting as well, but the approach is somewhat agnostic to how the changes actually get applied while still allowing you to enforce structure around what changes and how).

https://github.com/newhouseb/clownfish

Structural Alignment: Modifying Transformers (like GPT) to Follow a JSON Schema

Oh nice! I built a similar system a few weeks ago: https://github.com/newhouseb/clownfish

I think the main differentiating factor here is that this is better if you have a simpler JSON schema without enums or oneOf constraints. If you do have these constraints, i.e. let's say you wanted an array of different types that represented a items on a menu { kind: pizza, toppings: [pepperoni] } or { kind: ice_cream, flavor: vanilla | strawberry } then you would need something more sophisticated like clownfish that can ask the LLM to pick specific properties.

And on the other end, there's https://github.com/newhouseb/clownfish to force the model to produce structured output.

It doesn't help with repetition, but when it comes to force structure on the output data, this approach looks interesting:

https://github.com/newhouseb/clownfish

TL;DR: it exploits the fact that the model returns probabilities for all the possible following tokens to enforce a JSON schema on the output as it is produced, backtracking as needed.