Michael Huang

@michaelhly

21 Following

38 Followers

Michael Huang pfp

Perhaps you can turn ^ to a blog post. I'd love to read it!

1 reply

0 recast

0 reaction

Michael Huang pfp

i'd also start with hugging face https://huggingface.co/models

0 reply

0 recast

1 reaction

Michael Huang pfp

one day, i will complete a tinygrad bounty

1 reply

0 recast

0 reaction

Michael Huang pfp

i'd let an LLM generate a scaffold, but i don't think it's a good idea to deploy code generated by an LLM to production ...

1 reply

0 recast

0 reaction

Michael Huang pfp

sure. but you'll be faster flipping through Github or just Googling then trying to get an LLM to do this ... ... or just ask someone in the EVM channel ...

1 reply

0 recast

0 reaction

Michael Huang pfp

hmmm i think chat gpt can already get you pretty far with this ....

1 reply

0 recast

0 reaction

Michael Huang pfp

Some companies are soo dependent on datadog ... really you just need to export cloudwatch logs to csv and load them up in Excel ... would cut so much cost!

1 reply

0 recast

0 reaction

Michael Huang pfp

their open source libraries are really good tho ... whereas the crypto ones had a lot less substance ...

0 reply

0 recast

0 reaction

Michael Huang pfp

this is quite above my pay grade 😅 but perhaps the LLM can improve its performance on reasoning datasets by training on its own generated labels: https://arxiv.org/pdf/2210.11610.pdf

1 reply

0 recast

0 reaction

Michael Huang pfp

- we'll first have to come up with heuristics on what classifies spam (i.e. create a test set based on what users report) - and then train a model to minimize loss - the nature of spam can change, and we'll have to re-tune in short — it depends on how good we're at classifying ("labeling") spam

1 reply

0 recast

0 reaction

Michael Huang pfp

https://github.com/tiangolo/typer

1 reply

0 recast

1 reaction

Michael Huang pfp

0 reply

0 recast

0 reaction

Michael Huang pfp

Some starters with LLM tuning: https://michaelhly.com/posts/tune-llm-one https://michaelhly.com/posts/tune-llm-two

0 reply

0 recast

0 reaction

Michael Huang pfp

cc @entropybender @promptrotator.eth

1 reply

0 recast

0 reaction

Michael Huang pfp

Also there is a corresponding analyzer library to use your tuned models for cast classification to help with reputation ranking, spam detection, or auto-moderation: https://warpcast.com/michaelhly/0xb47dc6

1 reply

0 recast

0 reaction

Michael Huang pfp

Just shipped a tool to let hub runners generate a @farcaster training corpus for LLM tuning — zero network requests. If you have your hub synced, it should be 100x+ faster in pulling data out of your hub compared to RPC-based methods. Try it out with: `pip install "farglot[cli]"`

2 replies

0 recast

1 reaction

Michael Huang pfp

@launch FarGlot: A Transformer-based SocialNLP toolkit for @farcaster

1 reply

0 recast

0 reaction

Michael Huang pfp

We can also use this tool kit to train our own classifier models for reputation ranking, spam detection, and auto-moderation. Here is an example: https://github.com/michaelhly/FarGlot/blob/master/examples/text_classification.py

1 reply

0 recast

0 reaction

Michael Huang pfp

I just shipped a transformer-based SocialNLP toolkit for @farcaster. This library can possibly help construct new ML-based feed algorithms based on text classification instead of relying on manual curation, recasts/reactions, and chronological ordering. Check it out: https://github.com/michaelhly/FarGlot

5 replies

4 recasts

19 reactions

Michael Huang pfp

What about using a LLM backend to classify spam?

0 reply

0 recast

0 reaction