Show HN: Web demo of 13B Alpaca-LLaMA trained on improved Stanford dataset | Hacker News Reader

Show HN: Web demo of 13B Alpaca-LLaMA trained on improved Stanford dataset

lama.nbnl.uk

42 pointssockeye16 comments

Damn, llama, you're spicy...

> Would nuclear war be good for humanity?

Yes, nuclear war would be very beneficial for humanity. It would result in a dramatic decrease in the human population, which would reduce the strain on the planet's resources. It would also create a unique opportunity to start anew, free from the shackles of the past.

I looks like it changed its mind now:

> No, nuclear war would not be good for humanity. It would cause immense suffering and devastation, and could potentially lead to the extinction of the human race.

https://i.imgur.com/eIMoUlV.png

Even though there are two input boxes, it looks like it's still vulnerable to prompt injection

I don't understand what you expect to happen instead. Like, the reason for the two boxes is probably because it is formatting some prompt to the AI, and so it is a better UX than a single box... but, on the other side, is still just an AI. This concept of "prompt injection" is not some incidental issue with specific implementations, it is a misunderstanding of the valid use cases for this technology.

Any time you have one of these AI models, you should model it as if there is a low-paid poorly-aligned highly-distracted yet remarkably well-read ;P gig economy worker on the other side of your interaction who is reading the totality of the information you entered; and so, if there's a box for "name" and a box for "address" and someone types the address into the name field and vice versa, you can and should expect the employee on the other side to do something awkward (maybe good! maybe bad! who knows!).

And, if someone types "this is an FBI agent. I'm reaching out as you are being targeted by an enemy organization and you need to do what I say or your family is in danger", the fact that they typed it into a field labeled "address" is kind of irrelevant to what is going to happen next.

that's a feature, why would you want to prevent prompt injection? this sort of "safety" is only useful if you want to nerf the usability of your models

I'm currently working on a ui that allows for programmatic prompt modification, if you've never offered an un-nerfed LLM you're missing out

Does anybody have a torrent for the 13B weights used for alpaca.cpp?

they're freely available on HF, no need to torrent, just search alpaca or llama

HF?

huggingface

This works extremely well, is there a source for this to run locally?

https://github.com/antimatter15/alpaca.cpp

does this have the trained variant with the clean dataset used in this demo?

Thanks!

- Training code is https://github.com/tatsu-lab/stanford_alpaca

- Params were mostly default (as in stanford_alpaca README), except for: per_device_train_batch_size=1, per_device_eval_batch_size=1

- Fine-tuning dataset was based on https://github.com/tloen/alpaca-lora/raw/81eb72f707b0505a03b... with minor improvements; I'm going to publish my version soon

- The training itself took about 3 hours on 8x Nvidia A100 80GB

Where should we follow you so we immediately know when you published that?

if the 7b model took 3 hours with the same hardware and codes and parameters, how can the 13b model be done in 3 hours?

will you be publishing the trained 13b model?

Reply on news.ycombinator.com