this is interesting......should we explore adding a hackathon topic to investigate the use of models for nextflow coding ?

I don't get it tbh. isn't it yet another LLM playground?

it is but you can now pipe the model to existing CLI commands

so basically https://llm.datasette.io/en/stable ? (but honest opinion using a huge statistical model to get better logging, you are trying to solve the problem the wrong way around and in a really wasteful way)

true true....thanks for sharing that link though...didn't know abt it......there are just too many shiny new toys 😅

Has anyone fine-tuned an LLM using the nf-core (+ other Nextflow recommended) repos yet? I've thought about trying this, but definitely do not think I am the most qualified to give this a try 😅

I think @Florian Wuennemann played around with some

I am also definitely not the most qualified to have done so @Carson Miller 😂 ! When openAI announced custom GPTs I basically just wrote a prompt to force DSL2 for Nextflow code generation and pointed to the resources. I think the best approach for this type of thing would be proper RAG where one uses a good baseline model (chat-gpt4o or Claude or Llama3) and add all of the Nextflow, nf-core and seqera docs as some type of organized input that can be used to augment the knowledge of the model. As far as I understand it, fine-tuning requires a lot of high quality data and is a much bigger undertaking, whereas RAG basically just requires documents with knowledge not included in the model... 🤷