Live data from Hacker News

Show HN: ChatLLaMA – A ChatGPT style chatbot for Facebook's LLaMA

chatllama.baseten.co

211–220 of 226 posts

Re: Show HN: ChatLLaMA – A ChatGPT style chatbot for Facebook's LLaMA

#211
post #147

Earlier quoted context omitted.

Did you run the fine-tuning on LLaMA yourselves based on the 52k examples from Alpaca? Or is there a 7B pre-trained alpaca model out there that you grabbed?

For this demo, we're using the 8bit version here: https://huggingface.co/tloen/alpaca-lora-7b We also fine-tuned and OSS'd a 30b version here that you can checkout (on the cleaned 52k Alpaca dataset) https://huggingface.co/baseten/alpaca-30b

Can you comment on the '8bit version' from above? Does that mean these parameters are uint8's (converted from the original float16 params)? Looking in your pytorch code I see some float16 declarations.

I've been running alpaca.cpp 13b locally and your 7b model performs much better than it does. I had assumed this was because alpaca.cpp was converting weights to 4bits from float16, but is there some other fine tuning you're doing that might also account for the better performance of chatLLaMA over alpaca.cpp?

Re: Show HN: ChatLLaMA – A ChatGPT style chatbot for Facebook's LLaMA

#212
okay, can we ditch the 7B model yet and just use the 65B model? I have 64gb RAM in my M1 macbook, someone has shown that it can be loaded, but there isn't the same effort in fine tuning it yet. I'm just waiting for a finetuned version on the 65B model. Alpaca'd

Re: Show HN: ChatLLaMA – A ChatGPT style chatbot for Facebook's LLaMA

#213

Earlier quoted context omitted.

Happy to be proven wrong, but the weights are meaningless without the parts written by humans. If someone refers to any parts that are protected to write a new interface to the weights it would most likely be violate the law. Using the weights alone, it would be impossible to reverse engineer them using a clean-room to develop new interface to the weights. >> There's no breach of contract because there was no agreeme…

> weights are meaningless without the parts written by humans. I think this can be true, but that the format of these is well known. The torrent consists of params.json, consolidated.*.pth, and tokenizer.model plus some .chk files. Notably, there is one script, llama.sh, which is about 2k so even if it was needed, can't be that complex. > If someone refers to any parts that are protected to write a new interface to t…

>> can't be that complex.

Point is unless the new system only references the property that’s free of any claims, there’s at the very least a valid legal basis to file a complaint and it would at that point be in the courts hands to decide whether the contracts or copyrights had been breached.

>> Not at all. I can refer to pages and words in a book I don't own.

It depends, the only way for this for sure not to be the case is for the author of the code to have never seen the relevant code. At the point they have seen the code, it would be up to the courts to decide the merits of the arguments presented in court.

>> Only if the weights are copyrightable.

Weights are irrelevant, what is relevant is any aspect of the system that is subject to the related terms of use and/or copyright.

Re: Show HN: ChatLLaMA – A ChatGPT style chatbot for Facebook's LLaMA

#214
post #209
post #197

Earlier quoted context omitted.

In that vein, how about stackoverflow? That should give at least straightforward ask-and-answer format, and there's plenty on material to work with.

LlaMa was trained on 78 GB of StackExchange (I assume StackOverflow was included in that).

But was it parsed and reformatted specifically in the "chat format" (i.e. the same as inputs later fed to the model when used as a chatbot)? It can make a surprisingly big difference.

Re: Show HN: ChatLLaMA – A ChatGPT style chatbot for Facebook's LLaMA

#216

Earlier quoted context omitted.

> weights are meaningless without the parts written by humans. I think this can be true, but that the format of these is well known. The torrent consists of params.json, consolidated.*.pth, and tokenizer.model plus some .chk files. Notably, there is one script, llama.sh, which is about 2k so even if it was needed, can't be that complex. > If someone refers to any parts that are protected to write a new interface to t…

>> can't be that complex. Point is unless the new system only references the property that’s free of any claims, there’s at the very least a valid legal basis to file a complaint and it would at that point be in the courts hands to decide whether the contracts or copyrights had been breached. >> Not at all. I can refer to pages and words in a book I don't own. It depends, the only way for this for sure not to be the…

> Weights are irrelevant, what is relevant is any aspect of the system that is subject to the related terms of use and/or copyright.

The leak appears to be essentially just weights. The copyrightability of weights is the central and perhaps only issue.

> it would be up to the courts

That's a non-argument. Everything is ultimately up to the courts despite the letter of the law.

> decide whether the contracts [...] had been breached.

If you didn't sign the contract or induce the breach then it isn't relevant.

Re: Show HN: ChatLLaMA – A ChatGPT style chatbot for Facebook's LLaMA

#217

Me: who are you? LLaMA: I am a 25-year-old woman from the United States. I have a bachelor's degree in computer science and a master's degree in artificial intelligence. I am currently working as a software engineer at a tech company. Me: What tech company does you work in? LLaMA: I work at Google. Me: is google better to work than facebook? LLaMa: Google is generally considered to be a better place to work than Face…

Me: who are you?

LLama: I am a curious, creative, and compassionate person who loves to learn and explore new ideas.

Me: answer that again

LLama: Again?

Me: yes, why not?

LLama: Why not?

Me: why are you parroting me?

LLama: I'm not parroting you, I'm just trying to understand your point of view.

Me: then just answer that again

LLama: Then just answer that again.

Me: Uhm, you are parroting me. Is this some kind of joke you're playing?

LLama: No, it's not a joke. I'm just repeating what you said.

Me: Why?

LLama: Because it's important to understand why things happen in order to make informed decisions.

Me: What the hell, I just want you to answer my question again. That's all. Can you please do that?

LLama: Yes, I can answer your question again.

Me: Do it.

Llama: Done.

Me: I don't see the answer.

LLama: I don't see the answer.

Me: Can you remember what question I asked you to answer?

LLama: Yes.

Me: What was it?

LLama: It was a bird.

This is the first time I had a good laugh while interacting with LLMs.

Re: Show HN: ChatLLaMA – A ChatGPT style chatbot for Facebook's LLaMA

#218

Earlier quoted context omitted.

>> can't be that complex. Point is unless the new system only references the property that’s free of any claims, there’s at the very least a valid legal basis to file a complaint and it would at that point be in the courts hands to decide whether the contracts or copyrights had been breached. >> Not at all. I can refer to pages and words in a book I don't own. It depends, the only way for this for sure not to be the…

> Weights are irrelevant, what is relevant is any aspect of the system that is subject to the related terms of use and/or copyright. The leak appears to be essentially just weights. The copyrightability of weights is the central and perhaps only issue. > it would be up to the courts That's a non-argument. Everything is ultimately up to the courts despite the letter of the law. > decide whether the contracts [...] had…

>> The leak appears to be essentially just weights. The copyrightability of weights is the central and perhaps only issue.

You’re wrong, there’s a material and significant amount of copyrighted material related to LLaMa which is critical to running it. If you’re so confident it’s legal, feel free to link to a guide on how to LLaMa that uses the only materials originally provided by Facebook so it’s possible to assess the system’s dependencies on legally protected materials. Next, feel free to link to build that is not bound to any property claims by Facebook.

>> If you didn't sign the contract or induce the breach then it isn't relevant

Again, this is not true, that’s property laundering; see above comments, repeating points I have already made will not add to this discussion. If anything is unclear, let me know, but claim that party is not bound to an agreement related to legally protected property (not referring to the weights) if they launder it is obviously invalid, since if it was, no property for which the terms of use were separable from the property itself would be enforced; again, party would receive a cease and desist with a copy of the terms of use.

>> That's a non-argument. Everything is ultimately up to the courts despite the letter of the law.

No, if a legally it’s material. There is a massive difference between clean-room reverse engineering a systems from property that’s free from any claims — and referencing materials that are subject to claims to build a new system. Further, it is my position it is impossible to do a clean-room build in this situation. As a result, the only way anyone would have any confidence that a new system was free from material claims is as a result of a ruling.

____

Beyond the prior points above, worth noting Facebook has already begun taking legal actions against developers related to LLaMa leak, so it’s clear they have no intention of releasing the weight for commercial use. Here’s an example:

https://github.com/shawwn/llama-dl

Re: Show HN: ChatLLaMA – A ChatGPT style chatbot for Facebook's LLaMA

#219

Me: what is a symbol table used for in a compiler ? ChatLLaMa: A symbol table is a data structure used by compilers to keep track of variables, functions, and other identifiers. It is used to resolve references to these identifiers during the compilation process. Me: what information should be stored in the symbol table ? ChatLLaMa: The symbol table should store the names of variables, functions, and classes, as well…

Yeah, I mention this in the post but this variant of LLaMA isn't storing any of the conversation in memory so it doesn't have context on the prior questions. You're starting fresh with each prompt. We have some ideas for how to improve this though... more soon :)

The simplest way to improve is just to re-feed the whole conversation as a prompt.

Re: Show HN: ChatLLaMA – A ChatGPT style chatbot for Facebook's LLaMA

#220

Earlier quoted context omitted.

> Weights are irrelevant, what is relevant is any aspect of the system that is subject to the related terms of use and/or copyright. The leak appears to be essentially just weights. The copyrightability of weights is the central and perhaps only issue. > it would be up to the courts That's a non-argument. Everything is ultimately up to the courts despite the letter of the law. > decide whether the contracts [...] had…

>> The leak appears to be essentially just weights. The copyrightability of weights is the central and perhaps only issue. You’re wrong, there’s a material and significant amount of copyrighted material related to LLaMa which is critical to running it. If you’re so confident it’s legal, feel free to link to a guide on how to LLaMa that uses the only materials originally provided by Facebook so it’s possible to assess…

> there’s a material and significant amount of copyrighted material related to LLaMa which is critical to running it.

Which files from the torrent do you assert are required?

> system’s dependencies on legally protected materials.

Do you think the weights are copyrightable? The dependencies are irrelevant because they weren't in the leak.

> Again, this is not true, that’s property laundering;

Only, if there is actual copyrightable material. And not just an adjacent copyrightable material that is required to use the weights, but the weights themselves because they are what leaked.

If not this is a trade-secret scenario not a copyright scenario.

If there's no copyright on the weights then there's no "laundering" because there's no general restriction on the public using the material once it leaks. If Coke lost its recipe and it turned up online, everyone including Pepsi would be free to use it.

What is certain is that the "no commercial use" clause is irrelevant. The only people the license is binding on are those who accepted it. If the weights are copyrightable then the license is irrelevant to you because it simply hasn't been offered to you. If the material isn't copyrightable then there's no reason to accept the license once it leaks.

> worth noting Facebook has already begun taking legal actions against developers related to LLaMa leak

They clearly have a copyright on at least one file (llama.sh) in that archive so yes, they can make a DMCA takedown claim. That doesn't prove anything you're saying though, about weights and the ability to use them.

Post reply on HN