Building a database for LLM assistance with quest

[Diemethyl]

Im asking you all something that some of you may not really care for. some may call it 'cheating', but I am going to take the time to build a dataset for an LLM that could assist with monotonous tasks that are naturally part of development in quest.

in an ideal situation, the LLM could provide aslx snippets, and tell the user whereto place them.
the model would also be able to output scripts to meet the users desired functionality.

it is not uncommon for AI to come up with much simpler solutions to complex problems.
the model could take a load off when the user want to get to creating
it might help new users stay engaged

now:
it would hopefully be used to train a local model( or GPT if convenience is too much of a concern).

in the event of:
in the event GPT is used I will give out an open API key with US$10 per month for a year( US$ 120 dollars total) to the winner of a draw.

• ten dollars actually goes fairly far

to enter the draw

• users will need to give 5-10 examples (volume depends on quality) of asl or aslx along with a description of what it does in detail
  OR
• give one full game script in aslx with delimiters('---') and comments that explain the code in detail.

The draw would be held once 100 examples in total are received

for transparency, we will put those code samples here.
I will give the model, or the fully formatted dataset (depending on model ease of use) in this discussion at the time of training a model with 100 examples under the condition that the model can be fairly accurate.

[alexwarren]

I would be surprised if you could get people to contribute 100 examples here, especially as you're not offering very much in return. Is your plan to create a monetized custom GPT with this?

A lot of existing Quest knowledge has likely been ingested by the LLM providers already, but I think you would have better results by fine-tuning a model using what already exists - the forum here, the documentation and in Libraries and Code Samples in the old forum. It shouldn't "need" one big dedicated list of examples with explanations when the data already exists and the LLM can do a good job of parsing it all.

I was playing around with NotebookLM which has an impressively large context window - it will take up to 50 sources in the free version, each of which can be pretty large. You should be able to combine existing Quest docs and samples together into something that will fit, and it would be interesting to see what it can come up with.

[Diemethyl]

Sure, here's the corrected version of your text:

---

No, the idea of monetizing doesn't make sense for many reasons.

1. The dataset would be in this chat for all to see. All 3 of Quest's users.

2. You can't share a fine-tuned model without giving out a bunch of API keys. I know I could organize it in a way, but with that much effort, I may as well make my own dataset.

3. There's nothing to sell that wouldn't already be out in the open.

For the point about "nothing in return"? Well, a natural outcome of the data being posted right here kind of diminishes any monetary value.

However, you would get a dataset out in the open for all to use, plus I'd create a formatted version, plus a chance to get an API key for a year. I know $10 doesn't sound like much, but it goes very far for a casual user.

I've already made an Inform 7 dataset of over 3,000 examples. The problem with Quest is its small user base.

Then finally, monotonous tasks... I have arthritis, typing is painful for me. When I can ask AI to write up a function for me and get exactly what I need, it's great. On top of that, I am trying to make a very intricate RPG game with Quest, which may not be traditional, but I'm looking to procedurally generate content, which, no, doesn't mean that it will be generic, boring GPT-generated content. I need to lay out a framework. In fact, if it had enough examples a Quest user wanted, you could start giving it code examples to train it on the full Quest program and maybe even take Quest to its next version much quicker.

I'll be in the Inform 7 forums for now.

Note: No intent to insult, I honestly think not enough time was taken to think about what I previously said about transparency and where the entries would go.

[alexwarren]

Separate conversational tangent, but I'd be interested what "monotonous tasks that are naturally part of development in quest" means. That might point to ways we can make Quest itself better in future, rather than relying on AI to make things easier.

[Pertex]

I can't imagine any ‘monotonous tasks’, but it could help newcomers without programming experience to realise their ideas. However, I'm not familiar with LLMs and I wonder if it's so easy that newcomers can cope with it