[HN Gopher] StackLlama: A hands-on guide to train LlaMa with RLHF
       ___________________________________________________________________
        
       StackLlama: A hands-on guide to train LlaMa with RLHF
        
       Author : kashifr
       Score  : 119 points
       Date   : 2023-04-06 19:06 UTC (1 days ago)
        
 (HTM) web link (huggingface.co)
 (TXT) w3m dump (huggingface.co)
        
       | refulgentis wrote:
       | I've been on leave from work and hammering the GPT APIs since GPT
       | 3.5/ChatGPT was made available.
       | 
       | The local LLM stuff was a tad out of control from the drop, too
       | many people hand-waving about how they could get the 7B running
       | on a phone with quantization, but it was unintelligible, and not
       | "no-RLHF" unintelligible. Just FUBAR'd.
       | 
       | I tried the latest round of RLHF'd models yesterday, and I'm
       | officially publicly a skeptic now. These are an awful idea,
       | training on ShareGPT gets horrible results: I'm seeing it emit
       | the same exact answers ChatGPT does, but only a small fraction of
       | them.
       | 
       | I understand that it itself impressive for a certain crowd, and I
       | cede it's an accomplishment. However, it's an accomplishment that
       | enables no further accomplishment: using a stolen model to do
       | minimal RLHF that is really just overfitting on a subset of
       | answers from another AI. That's not RLHF at all. If it was, RLHF
       | isn't something you do in a weekend for $100, and pretty much
       | everyone outside OpenAI and Anthropic are learning that.
        
         | danielbln wrote:
         | How would you judge Open Assistant's approach?
        
           | refulgentis wrote:
           | I don't know much about it specifically but, heartily
           | endorse.
           | 
           | LAION was instrumental in early-ish AI art. I will always
           | cherish & remember when you had like 14 people in an IRC room
           | just playing around, fall/winter 2020. Now 3 of them have
           | companies around it, and the resources that were there to
           | enable ex. SD are similarly interested in LLMs.
           | 
           | This is excellent: open source is the way forward, just,
           | needs to be more coordination, expertise, and patience
           | involved. SNR ratio is way too low in general public spaces
           | like HN right now.
           | 
           | (I'm being exclusively negative, in recompense: we're at year
           | 1 of 100, the people wasting a ton of time replicating a
           | proof of concept crappy "RLHF" run and rushing to post are
           | learning too. The eternal golden summer starts now and
           | anything anyone is doing is helpful)
        
         | trifurcate wrote:
         | In my experience, the smaller models are almost completely
         | worthless as-is. 65B is the only decent one (I'd say just
         | behind gpt-3.5-turbo, and obviously it's not instruction tuned
         | but I mean the coherency of the core language model), and
         | understandably people aren't really paying attention or
         | devoting much resources to the largest one. 30B shows promise
         | for specific tasks with fine tuning, but 7B and 13B are just
         | toys.
        
       | kashifr wrote:
       | All the steps involved in training a LlaMa model to answer
       | questions on Stack Exchange data with RLHF.
        
         | ttul wrote:
         | You could of course use your own question and answer data to
         | refine the model using the same process. I wonder if anyone has
         | tried that yet to, for instance, fine tune LlaMa to answer
         | support queries for their company?
        
       | great_psy wrote:
       | Hopefully research like this will even out access to the new
       | tech. Maybe once we figure out a pretty good architecture we will
       | have something like chatBot.train(...) where we just feed some
       | data for the fine tuning.
        
       | scottydog51834 wrote:
       | I'd love a tool where I can upload a private dataset and RLHF a
       | model (even better if the tool provides the pre-trained model)
       | without me having to worry about GPUs, memory, commercial access,
       | or even writing any python code. I'd happy pay several hundred
       | and maybe several thousands for access to this.
        
         | chankstein38 wrote:
         | OpenAI allows you to fine tune any of their LLM models. I'm not
         | sure if your requirements were "without paying OpenAI
         | something" but if you are ok with that, fine tuning with them
         | is pretty straightforward. IANAL or anything but if I recall
         | they even promise they won't use your data for their training.
        
           | sebzim4500 wrote:
           | They let you do this with their GPT-3 derived models, but not
           | the GPT-3.5 or GPT-4 ones.
           | 
           | That would be unsafe lol
        
         | ttul wrote:
         | I think that tool would have broad appeal, but I wonder if the
         | most likely case is that it would be buried inside other,
         | higher-level systems, such as customer support automation SaaS.
        
         | Tepix wrote:
         | Even better would be a crowdfunding platform where you can
         | collect some money from a bunch of people and then use it to
         | fine-tune the bigger models (30B and 65B in the case of LLaMA).
        
       | pksebben wrote:
       | Glad to see more progress on open(ish) source versions. There's
       | so much more these things could do unfettered by corporate
       | motivations.
        
       | lumost wrote:
       | curious why all of these posts start with Llama vs one of the
       | many open source LLMs now. We have the Cerebrus releases,
       | Salesforce CodeGen-NL, and others.
        
       | Tepix wrote:
       | So, they are taking the Llama model released by Meta, doing a
       | little fine-tuning and then re-releasing the resulting model
       | under a different license?
       | 
       | That seems very sketchy. The Meta license grants a " _non-
       | exclusive, worldwide, non-transferable, non-sublicensable,
       | revocable, royalty free and limited license under Meta's
       | copyright interests to reproduce, distribute, and create
       | derivative works of the Software solely for your non-commercial
       | research purposes._ "
       | 
       | A better way would be to redistribute xdelta3 files so people
       | with access to the LLaMA model weights can use them to arrive at
       | the fine-tuned model weights. Or is there perhaps a better tool
       | than xdelta3 specifically for LLMs?
        
         | jimsimmons wrote:
         | HF wants to undercut OpenAI anyway possible.
         | 
         | My cynical take is that HF gives as much damn as OpenAI about
         | open source. It's just whatever gets you ahead of your peers.
         | 
         | Right now OpenAI has a massive advantage with GPT4 and their
         | RLHF stack. HF and maybe even Meta want to claw their way back
         | via crowdsourcing
        
           | refulgentis wrote:
           | This has ~0 to do with Hugging Face, Hugging Face is Github
           | for ML models
        
         | GaggiX wrote:
         | They only released the LoRA.
        
           | Tepix wrote:
           | Oh, you're absolutely right. I must have looked at the wrong
           | folder or something. Never mind then!
        
       | mcaledonensis wrote:
       | It is incapable of doing any arithmetic, e.g. on a question: _9 -
       | 4 =_                 Answer            There are a few other ways
       | to make this easier.            1. Keep the remainder as an
       | argument.            You can do that by rewriting your divmod()
       | function like   this:            def divmod(x, y):         return
       | x, (y % x)
        
         | sp332 wrote:
         | I asked a more verbose version of the same question, and it
         | started with a similar answer but added this:
         | 
         |  _[Edit]
         | 
         | In the comments, someone pointed out there were actually three
         | answers - one was 5; the other two being 1 and 2. Because these
         | numbers work out at the same value when they are multiplied by
         | 6, I have changed my answer to include all three
         | possibilities._
         | 
         | That was the best one I could get. It goes completely off the
         | rails even with the temperature quite low.
        
           | mcaledonensis wrote:
           | I'd call it a principle of invariance of compost piles.
           | Regardless of how long the compost pile is being stirred or
           | soaked, the product of the compost pile is compost.
        
         | drdaeman wrote:
         | It just generates some blabber that "seem" to relate.
         | 
         | I've asked it "How a raven is like a writing desk?" (assuming
         | that it's unlikely it was trained how to respond) and it just
         | started to "The answer can be found in Alice in the Wonderland"
         | then retell me the plot until it ran out of tokens. With a
         | lower temperature it switched to "Both are black" and something
         | about "dead men tell no tales".
         | 
         | I suppose trying to make an universalist model comparable to
         | GPT-3/4 with a drastically less number of parameters would
         | always produce subpar results, just because it can't store
         | enough knowledge. A specialist model, though, taught in depth
         | on one specific topic, may be still useful.
        
       | espadrine wrote:
       | One flaw that it has is to extend the question despite the prompt
       | requesting an answer.
       | 
       | For instance, I compare llama-7b-se-rl-peft to
       | OA_SFT_Llama_30B[0] on the following question: "What is the color
       | of the sky on Mars?"
       | 
       | llama-7b-se-rl-peft: "There are a few other related questions in
       | this site, but none addresses exactly what I'm looking for. If
       | you have been to mars (or if someone else has) can you please
       | tell me what the colour of the sky is and maybe how it looks from
       | different angles or at night time etc."
       | 
       | OA_SFT_Llama_30B: "The color of the sky on Mars depends on
       | various factors, such as time of day and year, atmospheric
       | conditions, and viewing angle. However, in general, the Martian
       | sky is often described as having a reddish or pink hue due to the
       | presence of iron oxide in its soil and atmosphere."
       | 
       | It could be the smaller size, or it could be the reward model not
       | incentivizing a proper understanding of the "Answer:" tokens.
       | Still, it is nice to see those open efforts.
       | 
       | [0]: https://open-assistant.io/chat
        
         | throwaway888abc wrote:
         | just tried same for curiosity "The Mars sky would be pink if
         | you were to look up and observe it in person. This is due to
         | the presence of finely dispersed iron oxide particles in its
         | atmosphere, which scatter blue light from the sun and cause the
         | sky to appear reddish-pink in color. This phenomenon is known
         | as "Rayleigh scattering." If seen through a telescope on Earth
         | however, the Martian sky may appear dark or even black
         | depending on atmospheric conditions and time of day/year. It
         | can sometimes be observed with a bluish tint during sunrise or
         | sunset when more sunlight reaches the planet's surface and
         | illuminates the dust particles in the air. During certain times
         | of year (such as spring and summer), a thin layer of water ice
         | clouds can also form at higher altitudes, causing the sky to
         | become slightly whiter than usual."
         | 
         | Model
         | 
         | OA_SFT_Llama_30B
        
       ___________________________________________________________________
       (page generated 2023-04-07 23:01 UTC)