outlines VS Constrained-Text-Genera

> [2]: https://github.com/outlines-dev/outlines?tab=readme-ov-file#...

It's interesting as speech recognition has become more popular than ever through services like Alexa, and other iot devices support for OS speech recognition

Unfortunately most implementations (especially those that are iot focused) don't have very important features for robust speech recognition.

1. Ability to enable and disable a grammar

[4] https://github.com/outlines-dev/outlines/tree/main

The next step is defining how to guide generation. For this step, we'll use the Outlines library. Outlines is a library for controlling how tokens are generated. It applies logic to enforce schemas, regular expressions and/or specific output formats such as JSON.

No benchmarks, just my anecdotal experience trying to get local LLM's to respond with JSON. The method above works for my use case nearly 100% of the time. Other things I've tried (e.g. `outlines`[0]) are really slow or don't work at all. Would love to hear what others have tried!

0 - https://github.com/outlines-dev/outlines

As I was playing with the Outlines library (https://outlines-dev.github.io/outlines/), I discussed with my friend Maxime how funny it would be if we set up a way to pair LLMs in chess matches till one wins. The first time I tried it, it required substantial prompt engineering to get some of those LLMs to propose valid moves. Large language models can mostly stay focused and even play rather well; see https://news.ycombinator.com/item?id=37616170 for example. However small language models aren't as easy to convince.

Some of those LLMs have seen very little chess notation and so after the first few opening moves there aren't any valid tactics, let alone strategy, so they would end up either repeating the same move, or hallucinate moves that are not valid (Kxe5, but there would be a queen on e5!)

Then Outlines came along and we could force them to pick valid moves with little cost! Maxime worked super fast and got a first version of this idea as a gradio space.

I think it is pretty fun to see the (mostly terrible, but otherwise valid) chess that those LLMs play. Maybe it will even be instructive to how we can create small LLMs that can play much better than the ones on the leaderboard.

Anyway, you can check it out here:

https://huggingface.co/spaces/mlabonne/chessllm

What is interactive about it: you can pick the LLMs from available models on HuggingFace (within reason, small LLMs are preferable so that the space does not crash) or push one of your own small models to HF and have it fight with others. At the end of the game the leaderboard is updated.

Hope you find it fun!

> The most obvious usage of this is forcing a model to output valid JSON

Isn't this something that Outlines [0], Guidance [1] and others [2] already solve much more elegantly?

0. https://github.com/outlines-dev/outlines

1. https://github.com/guidance-ai/guidance

2. https://github.com/sgl-project/sglang

And I think “just” Markdown doesn’t quite cut it for safe guidance. For example: directly generating content for your components. But I’m really excited about tooling like outlines appearing, with a greater focus on guided generation for structured data. Because this is often what we actually need!

It's pretty easy to force a locally running model to always output valid JSON: when it gives you probabilities for the next tokens, discard all tokens that would result in invalid JSON at that point (basically reverse parsing), and then apply the usual techniques to pick the completion only from the remaining tokens. You can even validate against a JSON schema that way, so long as it is simple enough.

There are a bunch of libraries for this already, e.g.: https://github.com/outlines-dev/outlines

https://github.com/outlines-dev/outlines/blob/7fae436345e621... squares with my experience using LLMs for anything real

  sequence = generator("Alice had 4 apples and Bob ate 2. Write an expression for Alice's apples:")

Oh my god. I wrote a whole library called "Constrained Text Generation Studio" where I mused that I wanted a "Photoshop for Text". I'm not even sure which work predates the other: https://github.com/Hellisotherpeople/Constrained-Text-Genera...

The core idea of a "photoshop for text", specifically a word processor made for prosumers supporting GenAI first class (i.e oobabooga but actually good) - is worth so much. If you're a VC reading this, chances are I want to talk to you to actually execute on the idea from the OP

I was working on this stuff before it was cool, so in the sense of the precursor to LLMs (and sometimes supporting LLMs still) I've built many things:

1. Games you can play with word2vec or related models (could be drop in replaced with sentence transformer). It's crazy that this is 5 years old now: https://github.com/Hellisotherpeople/Language-games

2. "Constrained Text Generation Studio" - A research project I wrote when I was trying to solve LLM's inability to follow syntactic, phonetic, or semantic constraints: https://github.com/Hellisotherpeople/Constrained-Text-Genera...

3. DebateKG - A bunch of "Semantic Knowledge Graphs" built on my pet debate evidence dataset (LLM backed embeddings indexes synchronized with a graphDB and a sqlDB via txtai). Can create compelling policy debate cases https://github.com/Hellisotherpeople/DebateKG

4. My failed attempt at a good extractive summarizer. My life work is dedicated to one day solving the problems I tried to fix with this project: https://github.com/Hellisotherpeople/CX_DB8

My mental model for LLMs was built by carefully studying the distribution of its output vocabulary at every time step.

There are tools that allow you to right click and see all possible continuations for an LLM like you would in a code IDE[1]. Seeing what this vocabulary is[2] and how trivial modifications to the prompt can impact probabilities will do a lot for improving the mental model of how LLM operate.

Shameless self plug, but software which can do what I am describing is here, and it's worth noting that it ended up as peer reviewed research.

[1] https://github.com/Hellisotherpeople/Constrained-Text-Genera...

Yeah, the LLM outputs a distribution of likely next tokens. It is up to the decoder to select one, and it can use a grammar to enforce certain rules on the output. https://github.com/Hellisotherpeople/Constrained-Text-Genera... or https://github.com/ggerganov/llama.cpp/blob/master/grammars/... for example.

I am in love with this, I tried my hand at building a Constrained Text Generation Studio (https://github.com/Hellisotherpeople/Constrained-Text-Genera...), and got published at COLING 2022 for my paper on it (https://paperswithcode.com/paper/most-language-models-can-be...), but I always knew that something like this or the related idea enumerated in this paper: https://arxiv.org/abs/2306.03081 was the way to go.

I agree with you, and I'm SHOCKED at how little work there actually is in phonetics within the NLP community. Consider that most of the phonetic tools that I am using to enforce rhyming or similar syntactic constrained in constrained text generation studio (https://github.com/Hellisotherpeople/Constrained-Text-Genera...) were built circa 2014, such as the CMU rhyming dictionary. In most cases, I could not find better modern implementations of these tools.

I did learn an awful lot about phonetic representations and matching algorithms. Things like "soundex" and "double metaphone" now make sense to me and are fascinating to read about.

https://github.com/hellisotherpeople/constrained-text-genera...

Just ban the damn tokens and try again. I wish that folks had more intuition around tokenization, and why LLMs struggle to follow syntactic, lexical, or phonetic constraints.

My work on constrained text generation / filter assisted decoding for LLMS is cited in this article! One of my proudest moments was being noticed by my senpai Gwern!

https://paperswithcode.com/paper/most-language-models-can-be...

I want to update that just because GPT-4 appears to be far better at following constraints, doesn't mean that it's anywhere near perfect at following them. It's better now at my easy example of "ban the letter e" but if you ask for several constraints, or mixing lexical and phonetic constraints, it gets pretty awful pretty quickly. Filter assisted decoding can make any LLM (no matter how awful they are) follow constraints perfectly.

I can't wait to get someone whose better at coding than me to implement these techniques in the major LLM frontends (oogabooga, llamma.ccp, etc) since my attempt at it was quite poopy research code: https://github.com/hellisotherpeople/constrained-text-genera...

The paper at COLING 2022 that I wrote, titled "Most language models can be poets too" included a GUI constrained text generation studio that I market as being "Like Photoshop but for text"

https://github.com/Hellisotherpeople/Constrained-Text-Genera...