rnosov

The report cites both GPT-3.5 and GPT-4 scores on page 7 [1]. I've checked the numbers and they compare FreeWilly2 to GPT-3.5. For example, HellaSwag score of 85.5% corresponds to GPT-3.5.

[1] https://arxiv.org/pdf/2303.08774v3.pdf

we are training a full suite of models, with the first becoming available in the coming weeks.

Sounds like they already have the compute and began training.

For example, you can look up semantically relevant ( to a user query ) paragraphs from some internal document. Then, include them in the LLM context so it would know how to answer the query. Basically, that's the idea behind many ChatGPT plugins.

row 7 is the name of the dataset, you might need to load it yourself

I'm going through the dataset with your datasette tool and it looks like it might be a good idea to clean things up a bit. There are many duplicates[1], creepypastas[2] and other strange things in there.

[1] https://lite.datasette.io/?json=https%3A%2F%2Fraw.githubuser...

[2] https://lite.datasette.io/?json=https://github.com/databrick...

EDIT: Maybe I'm passing link wrong, the query I'm using is

select count(instruction), instruction, group_concat(context, ' ============= ') as c, group_concat(response, ' ============= ') as r, group_concat(category, ' ============= ') as cat from [databricks-dolly-15k] group by instruction having count(instruction)>1 order by count(instruction)desc limit 100

[databricks-dolly-15k] should be the name of dataset, first column is the number of instruction duplicates

Creepypastas are responses to instruction:

Imagine you are the last person on Earth. Write a diary entry describing your thoughts and feelings.

They are leveraging Apple’s Metal Performance Shaders[1] not the neural engine. From the chart, it looks like you might get ~20x max boost on inference over plain CPU. Obviously, it's not like having RTX 4090 but better than nothing.

[1] https://pytorch.org/blog/introducing-accelerated-pytorch-tra...

It looks like you're on python 3.11 which has some issues with Pytorch. Downgrade to python 3.10 and try running it again.

A question to the author. Can you perform an ablation study with respect to the chunks? In other words, if you put in the context irrelevant/random chunks from the document would the quality of answers decrease/stay similar?

Potential issue might be that chunks just serve to activate massive knowledge of GPT4 and not actually used as a basis for an answer. For example, GPT4 has surely seen Dune in its training corpus and could be answering from memory.

1) I've looked at both codebases and this one is definitely a derivative of the nanoGPT. You can compare all three implementations yourself as they are actually surprisingly compact and readable.

2) The issue whether weights are copyrightable at all has not been settled yet. If they are, there is a fair use doctrine that allows transformative works of a copyrighted work. The line is a bit blurry but consider Cariou v. Prince case[1] where addition of colour to some black and white photos was considered enough to be transformative. Similarly, full fine tuning on current news or adding visual modality could potentially create a brand new model in the eyes of a law.

[1] https://cyber.harvard.edu/people/tfisher/cx/2013_Cariou.pdf

I might be missing something but it looks to me that actually running this "open" model requires special hardware only accessible with a cloud subscription with 60 000 USD / week minimum spend[1]. Can anyone confirm if you can run it on your own hardware? If software is open but hardware is locked I don't see the point.

[1] https://www.hpcwire.com/2021/09/16/cerebras-wafer-scale-engi....

EDIT: Ok, looks like I've missed the hugging face repo. The language they use is a bit confusing.

The FTC link has an example of the only newspaper in town refusing to deal with customers who are also running ads on a radio station. Do you think if the newspaper dressed such refusal as a ToS violation it would fly with FTC?

Google might be banning people for enforceable violations of their ToS but imagine the uproar if they banned a Bing engineer for using Google search to find solutions for some Bing problem (which is similar to the problem here). The upside for Google or OpenAI would be somewhat limited but the downside is almost boundless.

That would be anticompetitive practice that is actually against the law in many countries[1]. In the unlikely event of OpenAI ever engaging in such things they will be sued into oblivion.

[1] https://en.wikipedia.org/wiki/Refusal_to_deal

ToS are not the law. It would be similar to your power company claiming copyright over the code written using "their" electricity. Not going to happen. I wouldn't be too concerned.

They are instruction tuning it using the dataset released by stanford-alpaca team. The dataset itself is synthetic (created using GPT-3) and somewhat noisy and in my view can be easily recreated if OpenAI ever tries to go after it (which is very unlikely). Anyway, facebook has nothing to do with anything used by this project.

This has nothing to do with facebook. The foundational model here is GPT-J which is opensource and safe to use. Sadly, it is inferior to state-of-the-art models such as LLaMA.

This is a really troubling development. Anyway, kudos to you for standing up to the man.

There are benchmarks in the original LLaMA paper[1]. Specifically, on page 4 LLaMA 13B seems to beat GPT-3 in BoolQ, HellaSwag, WinoGrande, ARC-e and ARC-c benchmarks (not by much though). Examples that you've seen are likely to be based on some form quantisation / poor prompt that degrade output. My understanding that the only quantisation that doesn't seem to hurt the output is llm.int8 by Tim Dettmers. You should be able to run LLaMA 13B (8 bit quantised) on the 3090 or 4090 consumer grade GPU as of now. Also, you'd need a prompt such as LLaMA precise[2] in order to get ChatGPT like output.

[1] https://arxiv.org/pdf/2302.13971v1.pdf

[2] https://www.reddit.com/r/LocalLLaMA/comments/11tkp8j/comment...

You describe supervisor approach as:

One common suggestion is to have another LLM look at the input intently with the instruction to determine whether it is malicious.

Preflight prompt check is actually opposite of that in a sense that it is more like a concurrent injection. You embed a random instruction with a known output and compare completions. As far as I know, nobody has been able to bypass it so far. False positives would be a problem but as you point out microsoft has no issue with collateral damage and blocking all github subdomains wholesale at the moment.

Similarly, you can embed a second instruction during preflight check asking for a count of [system] mentions. Since you know this number beforehand, if it changes it will signal that the prompt is poisoned.

Quite an interesting article. The Vice example is hilarious. But for all doom and gloom you haven't addressed the most obvious mitigation - Preflight Prompt Check [1]. It would be trivial to detect toxic prompts and halt further injection. Surely there will be other mitigations to follow.

[1] https://research.nccgroup.com/2022/12/05/exploring-prompt-in...

There is a big AI division of Google located in the UK (DeepMind). Could be related to that.

https://arxiv.org/abs/2210.17323

I've read the paper and to be honest I'm not sure what to make of it. Their headline benchmark is perplexity on WikiText2 which would not be particularly relevant to most users. If you look at the tables in the appendix A.4 with some more relevant benchmarks you'll sometimes find that straight RTN 4 bit quantisation beats both GPTQ and even full 16 bit original! No explanation of it is given in the paper.

I don't think that the model they are using is private. You can download it[1] and run yourself. Looks like it is GPT-NeoX which was fine tuned using the open source LAION OIG dataset[2].

[1] https://huggingface.co/togethercomputer/GPT-NeoXT-Chat-Base-...

[2] https://huggingface.co/datasets/laion/OIG

You might want to tune the sampler. For example, set it to a lower temperature. Also, the 4bit RTN quantisation seems to be messing up the model. Perhaps, the GPTQ quantisation will be much better.

The post has a workaround for the PyTorch issue with Python 3.11. If you follow the repo instructions it will give you some rather strange looking errors.

Seconded. Do write it up.

I see vast.ai listing interruptible instance with a single A100 80GB at $1/hour which is pretty reasonable. ChatGPT plus is $20/month which would be roughly 20 hours of use and I wont't be lectured like I'm in a kindergarten or something.

A bonus point would be to make the writeup accessible for AI challenged developers. Asking for a friend.

Hmmm, the Github repo suggests that you might be able to run the 65B model on a single A100 80gb card. At the moment, the spot price on Google cloud for this card is $1.25/hour which makes it not so crazy expensive...

You can read the original LLaMA paper which is pretty accessible[1]. For example, they claim to outperform GPT-3 on HellaSwag benchmark ( finishing sentences ). You can find examples of unfinished sentences in the HellaSwag paper [2] on page 13. Unfortunately for LLaMA, most people would be probably just asking questions about Captain Picard and so on, and on this benchmark LLaMA significantly underperforms compared to OpenAI models (thats's from their paper).

[1] https://research.facebook.com/file/1574548786327032/LLaMA--O...

[2] https://arxiv.org/pdf/1905.07830.pdf

They forgot trailing zero. It should be 130 sq ft[1]. Looks like no window too.

[1] https://www.lifeatseacruises.com/pricing-and-cabins

There are plenty of places on dry land that are a LOT cheaper than 30k/year. Probably would be more comfortable too.

I'm not entirely sure what sort of courses you are planning to offer, but what I can tell you is that course material quality at traditional higher education providers (bottom tier) is appalling (and they know it). If you can get it right, that would be worth a lot more than $1m/year. At the moment, the most obvious bottleneck would be the relatively small context window of GPT models (<3k words). It is speculated that OpenAI plans to increase it to 24k words, which would make it a lot more feasible though.