data
stringlengths 115
7.61k
|
|---|
SteveMcQueen#3713: Joined the server.
Sid#2121: Joined the server.
Sid#2121: 👋
Daj#7482: Hello hello
Daj#7482: I'll get myself up to speed on how to properly set up and admin a discord server soon
Sid#2121: wish i could be of any help at all in that regard lol
Daj#7482: I think our needs are pretty minimal haha, but feel free to request channels or other things we might need
Sid#2121: can we get a #datascripts channel or something? to collect all our individual scripts for data gathering
Daj#7482: Sure and we can use #the-pile for discussion?
Sid#2121: yeah
Sid#2121: and maybe separate chat channels from link/resources channels, if you can do that
Daj#7482: Ah yes this is all pretty intuitive
Daj#7482: Do we need/want roles and permissions and stuff?
bmk#1476: Joined the server.
Daj#7482: Hey hey
Sid#2121: i wouldn't really know what to do with a role anyway lol
Daj#7482: Just setting things up, please request any channels/roles/settings you want
bmk#1476: let's make an info channel where i can dump all of my documents about the project
bmk#1476: can haz channel adding perm
Daj#7482: Sure let me quickly figure out how to do that |
bmk#1476: https://cdn.discordapp.com/attachments/729741769738158194/730093559427891250/unknown.png
bmk#1476: wrote this up
Daj#7482: Try that
Daj#7482: I do like the colors lol
Sid#2121: ooh how do i get a fancy colour
Sid#2121: DAMN
Sid#2121: already have one
Sid#2121: nice
Daj#7482: Yea I just created a GPT Neo role and gave it to you all
Daj#7482: You have most permissions but I don't think we need to worry about trolling
bmk#1476: that reminds we i think we should choose a better name than gptneo
Sid#2121: aw i kinda like it
Sid#2121: but shoot
bmk#1476: 1. it makes it sound like we have something to do with openai
Daj#7482: I kinda like it too, but open to suggestions
Daj#7482: (we are using like 70% OA code and copying their papers lol)
Daj#7482: but yeah I see what you mean
bmk#1476: like people might think we're affiliated
Daj#7482: Eh I doubt that'd be a problem
bmk#1476: I'd vote for something with Libre or Open or sth in the name |
Daj#7482: but if you have a cooler name totally open
bmk#1476: LibreLM
bmk#1476: NeoLM
Sid#2121: I like using GPT because, well, it's gpt
Sid#2121: we're not doing anything totally novel
Sid#2121: aside from open sourcing it
bmk#1476: LibreGPT
Daj#7482: I feel like we might not wanna overuse the Libre/Open names
Daj#7482: (like OpenAI does 🔥)
Sid#2121: https://tenor.com/view/rap-rap-battle-gif-5518154
Daj#7482: We can still change it before we release to the public
Daj#7482: So we can think on it
bmk#1476: yeah i guess
bmk#1476: just voicing my opinion rn so it's not a surprise later
Sid#2121: i'm not totally attached to neo, but i feel like keeping the gpt is important
Daj#7482: Will keep in mind for sure
Daj#7482: btw we do have a channel we could use instead if we wanna try and be organized hah #gpt-neox-devs
Daj#7482: Name can still be changed down the line easily
SteveMcQueen#3713: Random question, but how do you guys plan on training this absolute beast of a model? I mean unless someone has basically unlimited credits laying around
Daj#7482: I have basically unlimited credits laying around |
SteveMcQueen#3713: Well that solves that
Daj#7482: Yea haha, I was an early adopter of TFRC and know the guys in charge semi personally
Daj#7482: So we have theoretical access to up to 2048 TPU cores whenever
Daj#7482: (though that many are usually not free)
Daj#7482: I made a #links channel meant for clean posts of links to papers with at most a paragraph of description, and a more general #deleted-channel channel for discussing papers and such. Sound reasonable?
bmk#1476: yeah
Daj#7482: I also think we could rename #data-sources to data-sources
bmk#1476: sure
Daj#7482: Nice everything seems so orderly now
bmk#1476: also i'm putting a star next to important datasets
bmk#1476: so we can prioritize
bmk#1476: any idea how to download from aws datasets?
Daj#7482: Uhhh
Daj#7482: nope
Daj#7482: It seems like it may require opening the dataset on an AWS instance or something
Daj#7482: Which would be a hassle but not a big roadblock
bmk#1476: it's also not text, but rather a knowledge base
bmk#1476: so lots of commonsense info in a not-LM friendly format
Daj#7482: Super interesting but we'd have to put in some thought as to how/if we use it
Daj#7482: Procedurally generate true sentences from the knowledge? |
Daj#7482: That sounds cool
bmk#1476: yeah
bmk#1476: also once we have this dataset and we've trained a GPT3 replica we should publish a paper about it, with the focus being on the dataset
Daj#7482: You think we could do that? I'd absolutely love to
bmk#1476: it's worth a shot
Daj#7482: Though I don't have the connections to get it on arxiv or published personally
bmk#1476: getting it on arxiv should be trivial
Daj#7482: But I've written/edited papers before so I think I can be helpful
Daj#7482: You need an endorsement for arxiv iirc
bmk#1476: not too hard to get
Daj#7482: haha well I will defer to you there
bmk#1476: I can ask around, at least
bmk#1476: hopefully that works
Noa Nabeshima#0290: Joined the server.
Daj#7482: Would be awesome, I'd love to publish
Daj#7482: Hey Noa, server is an early stage. Resources channels are for links and no chat, rest is free range
Daj#7482: You should have permissions for most things
bmk#1476: any other directions for data that we don't have?
Daj#7482: I'd add links to what we're already working on perhaps? (CC, WT, C4 for the cleaning methods, wikipedia...)
Daj#7482: Maybe also libgen? Or should we have like a seperate pirate channel? |
bmk#1476: I'll add a "current datasources" section
bmk#1476: to the document
Daj#7482: Can you pin your gdoc to the gpt-neo channel?
bmk#1476: done
Daj#7482: Lovely!
bmk#1476: https://cdn.discordapp.com/attachments/729741769738158194/730105119252807782/unknown.png
bmk#1476: am i missing anything important
Daj#7482: Seems good to me, though I may be forgetting some thing or the other we had discussed at some point
bmk#1476: and of these, wikipedia, opencaptions and bookcorpus are done, CC, CORE, and libgen are in progress, and WT2, irc are to do
Daj#7482: Nice
Daj#7482: We're really making nice progress
bmk#1476: i mean CC core and libgen are by far the biggest pita
Daj#7482: Of course yea
bmk#1476: and also sadly the most important
bmk#1476: we can make a mini version of the dataset though
Daj#7482: but still I'm just surprised this didn't peter out within 24 hours
bmk#1476: yeah me too
bmk#1476: WT1 + BC + Wiki + OC + IRC + small preliminary pieces of LG and CC and CORE
bmk#1476: and we can train a slightly larger GPT2 with it
Daj#7482: Sure that shouldn't be a problem |
Daj#7482: _Larger_ GPT2 won't work without TFM
Daj#7482: but we could experiment with 1.5B right away basically. Just need to converto tfrecords and upload to google cloud
bmk#1476: i have no idea how to do that
Daj#7482: I can write the scripts for it if we have the texts
Daj#7482: Not that hard
bmk#1476: what's more worth doing, GPT2+more data or charGPT2&same data
bmk#1476: I'm willing to help do some more work on the TFM front to help make it happen
Daj#7482: I think getting a GPT2 to OA levels of performance would be an important proof of concept
bmk#1476: ok
Daj#7482: 1.5B can be trained within a week
Daj#7482: And if we can get the right performance I'd be confident in scaling up
bmk#1476: will throwing more data at GPT2 make it better?
Daj#7482: Probably
bmk#1476: i thought the limiter here was model capacity
Daj#7482: If it's high quality
Daj#7482: Well yes
Daj#7482: You might be right
Daj#7482: This is an empirical science haha
bmk#1476: haha
Daj#7482: Which is why it's worth doing these experiments before going the full length |
bmk#1476: ok well we have WT1, BC, Wiki, OC so far
bmk#1476: I can make some LG from the data i already have downloaded
bmk#1476: and is the core download done?
Daj#7482: You can see my old tfrecords scripts in the GPTNeo/datasets folder if you wanna have a look. Though they produce the "sample randomly" kind of tfrecords, but GPT3 linearly reads through all the data
Daj#7482: I think goolu was doing CORE?
bmk#1476: I think I'll just get everything into either compressed json list or tgz of random files format
Daj#7482: Basically we just need to be able to ingest a stream of text documents that we can glue together with <|endoftext|> and encode into BPE or whatever other encoding we're using
bmk#1476: I can do that
Daj#7482: The final tfrecords samples are just chunks of integers after all
bmk#1476: I'll get together a script that takes either of my common formats and provides a python iterator
Daj#7482: Awesome
bmk#1476: is there a way to weight part of the data more?
Daj#7482: Yes
bmk#1476: just return multiple copies?
Daj#7482: If they are in seperate files my GPT2 input code can weight collections of files differently
bmk#1476: alright I'll provide a `random data file -> [String]` converter
Daj#7482: Sounds good
Daj#7482: Will also be trivial to parallelize just by ingesting different files
bmk#1476: re: libgen&bookcorpus data: is it a problem if we have a block of copyright info and stuff at the beginning of each file
bmk#1476: ```Also by Jim Lacey |
Takedown: The 3rd Infantry Division’s
Twenty-one Day Assault on Baghdad
Pershing: A Biography
The Canons of Jihad: Terrorists’ Strategy for Defeating America (editor)
The Making of Peace: Rulers, States, and the Aftermath of War (co-editor, with Williamson Murray)
Copyright © 2011 by James Lacey
Maps copyright © 2011 by Jeffrey L. Ward
All rights reserved. |
Published in the United States by Bantam Books, an imprint of The Random House Publishing Group, a division of Random House, Inc., New York.
BANTAM BOOKS and the rooster colophon are registered trademarks of Random House, Inc.
Grateful acknowledgment is made to Pantheon Books, a division of Random House, Inc., for permission to reprint excerpts from The Landmark Herodotus: The Histories by Robert B. Strassler, translated by Andrea L. Purvis, copyright © 2007 by Robert B. Strassler. Reprinted by permission of Pantheon Books, a division of Random House, Inc.
LIBRARY OF CONGRESS CATALOGING-IN-PUBLICATION DATA
Lacey, Jim
The first clash: the miraculous Greek victory at Marathon and its impact on Western civilization / Jim Lacey.
p. cm.
eISBN: 978-0-553-90812-1
1. Marathon, Battle of, Greece, 409 B.C.—Influence. 2. Greece—Civilization—To 146 B.C. 3. Iran—History—To 640 A.D. I. Title.
DF225.4.L33 2011 938′.03—dc22 2010046214 |
www.bantamdell.com
Jacket design: Christopher Sergio
Jacket photograph: © National Museum of Scotland / The Bridgeman Art Library
v3.1
```
bmk#1476: these files are a few hundred kb-1mb anyways so this is a tiny fraction of the data
Daj#7482: It's not breaking or anything
Daj#7482: Dunno if it's worth investing effort to clean it
Daj#7482: Though it wouldn't hurt to clean it ofc
bmk#1476: eh i'll just leave it
Daj#7482: Yea probably fine
bmk#1476: https://gist.github.com/leogao2/dfee376d02b7119ebc0770758e5c19bb
SteveMcQueen#3713: Would a simple solution not be to just cut out the first 1-2 pages?
bmk#1476: analysis of unicode occurring in filtered CC
bmk#1476: i guess but then we lose useful data too |
Daj#7482: It might actually be on net worth it to cut out the first page or so. Dunno something we can think about/experiment with down the line
bmk#1476: ok ill do cutting
Daj#7482: This is the kind of nitty gritty trial-by-error data people will benefit from later too
Daj#7482: lol we could do like a statistical test how likely it is that cutting pages deletes copyright or real data using a spot sample (yes I'm studying for my stats exam atm)
bmk#1476: or should i just give you the raw uncut data
SteveMcQueen#3713: hahaha my thoughts exactly
Daj#7482: and I think we'll train with BPE first before we can do the longer ctxs
SteveMcQueen#3713: random sample of around 1000 books and see whats on the first apge
Daj#7482: Shouldn't be too hard to do
bmk#1476: one issue: libgen dumps are not homogenous
bmk#1476: they vary a lot from dump to dump
bmk#1476: and this particular dataset is from a very specific subset of dumps
Daj#7482: Yea but a totally subjective look at data is often totally sufficient
Daj#7482: Just do what you think sounds sensible
SteveMcQueen#3713: Ah yeah fair point, technically you can just download x different data sets shuffle them and take the top 1k or so
bmk#1476: like, it's not a representative sample
Daj#7482: Fair
SteveMcQueen#3713: but get that it might be not really worth it
Daj#7482: We'll just keep iterating and experimenting
bmk#1476: anyways i'll just give you the raw untouched data |
Daj#7482: Data science and ML is black magic™️
bmk#1476: for SmallLibGen™️
Daj#7482: Which is imo a technical term that means "needs a lot of subjective judgement"
Daj#7482: Sounds good
bmk#1476: should we switch from json to bson for extra speed
bmk#1476: (with backwards compat)
Daj#7482: Any speed gain is worth it at these sizes
bmk#1476: ok
bmk#1476: really flooring it on my computer https://cdn.discordapp.com/attachments/729741769738158194/730122885863178280/unknown.png
bmk#1476: > total received: 962.0GB
goolulusaurs#1571: Joined the server.
Sid#2121: if we're going to be including IRC chatlogs we might need to *seriously* think about bias https://cdn.discordapp.com/attachments/729741769738158194/730173220971347998/Screenshot_2020-07-07_at_23.26.51.png
Daj#7482: Good point, IRC and chatlogs are probably cesspools
Daj#7482: Of course, that might add to the cyberpunk aesthetic. But I agree this is a non trivial issue, I'll think about it some more
Sid#2121: i mean, idk if CC is any better? I actually have no clue how they scrape
Sid#2121: or where from
Daj#7482: Ctrl-F "n-word" should answer our questions
Sid#2121: yes
Daj#7482: speaking of...4chan has full archives floating around
Sid#2121: oh god no |
Daj#7482: Just saying
Daj#7482: Haha
Sid#2121: I actually think I might rather not, idk
Daj#7482: Nah I getcha
Sid#2121: it's something to think about
Daj#7482: I just have a twisted sense of humor
Daj#7482: maybe we'll make GPT-Toxic
Daj#7482: as a side project
Sid#2121: lmao. just entirely 4chan?
Sid#2121: 4chan distilled into a hyperintelligent ai ono
Daj#7482: I'm not saying it's a good idea
Sid#2121: it's *an* idea
Daj#7482: But I am saying it's a _hilarious_ idea
Daj#7482: We could unleash it as a chat bot onto 4chan and see if anybody notices
Sid#2121: omg
Sid#2121: i want to create it now entirely for that purpose
Daj#7482: We really are in good company here haha
Daj#7482: Add it to the bucket list
Daj#7482: Can be a 1.5B variant
Sid#2121: i mean, collecting the data can't hurt and we can always finetune the smaller models on it if we don't end up using it, yeah |
Daj#7482: Agreed
Daj#7482: 4chan archives keep popping up and dying all the time, so we might have to look around a bit
bmk#1476: 4ch archives are really fragmented unfortunately
Daj#7482: Still, if no one else wants to I would (when I'm more free) try to collect as much as I can lol
Sid#2121: who runs 4ch these days?
Daj#7482: I have no idea, I haven't been on there in like 5 years
Daj#7482: >Google 4chan archives
>"'My Little Pony' Fans Confront Their Nazi Problem - The Atlantic"
Daj#7482: kek
Sid#2121: i want 8ch archives too just for all the qanon madness
Daj#7482: Oh boy
Daj#7482: Yes
Daj#7482: ~~Is this abusive towards the AI?~~
Sid#2121: but then we're really getting into some godawful degenerate shit
bmk#1476: how about no
Daj#7482: Fair
Daj#7482: I'll just put it on the sidelines...for now
bmk#1476: i mean im ok with data that may have a high level of toxicity
Daj#7482: We definitely won't include it in any main projects
Sid#2121: yeah ok 8ch can stay out |
Daj#7482: Since it's also probably shit
bmk#1476: but 4chan isnt just "high level of toxicity", it's literally the definition of toxicity on the internet
Daj#7482: I know worse places tbh
bmk#1476: on the popular internet
Sid#2121: well, 8ch
Daj#7482: it's more just 14-18 year olds being edgy
bmk#1476: *ch
Daj#7482: But yeah I vote don#t include it in the GPT3 or 1T runs
bmk#1476: yeah that's the most toxic demographic on the internet
Daj#7482: But I might privately play with it for shit and giggles
Sid#2121: > it's more just 14-18 year olds being edgy
@Daj it's not even just that though. I wish it were that simple
bmk#1476: yeah that's all fine and dandy
Daj#7482: Unfortunately I know that's at least partially true Sid
Daj#7482: I used to be a pretty active user when I was 17-18 lol
Sid#2121: I never could get past the nazis
Sid#2121: videos of ppl dying, strange porn, sure
Sid#2121: but yeah. always felt unwelcome
Daj#7482: Really? That's what did it for you? lol I get you
Daj#7482: Oh yeah it's a shit place I'm not defending it |
bmk#1476: i think we should get the more important data, like WT2, before turning to such.. unorthodox sources
Daj#7482: I hate everyone I met there lol
Daj#7482: Absolutely
Sid#2121: @bmk how is the AWS link you posted to the enron data 210Gb lol? a source I've found is like 400mb
Sid#2121: how is that formatted
bmk#1476: attachments
Sid#2121: right. don't really need them
bmk#1476: ye
SteveMcQueen#3713: btw even though CC might have the reddit data pushshift has a full dump by month of all comments/submissions (https://files.pushshift.io/reddit/comments/)
SteveMcQueen#3713: It is not fully updated (missing like 6 months or so), and it comes with a decent bit of metadata that needs to be removed, but otherwise should be a significant amount of pure text there
Sid#2121: the more the better. We'll be deduping anyway
Sid#2121: post it up in #the-pile
Sid#2121: actually no lol
Sid#2121: #data-sources ?
SteveMcQueen#3713: Yeah probably fits best there
SteveMcQueen#3713: Quick question on methodology for creating our corpuses (might have missed this), but what thoughts do we have in regards to meta tagging things? Will it just be one huge ass dump and then except the model to shift through it by its lonesome or are we planning on tagging/ordering things ahead of time? Simple example could be Reddit data, where the order and structure of the comments are important (e.g. what is nested under which submission/comment chain)
SteveMcQueen#3713: And/or any idea how OA tackled this?
Sid#2121: it's just a huge dump of text, non-text data like structure is superflous and can't teach the model anything
Sid#2121: obviously separate sources will be kept separate, it won't just be one massive 10TB text file but
Sid#2121: we essentially want to parse out as much non-text as possible |
Sid#2121: I mean, the order is important, but I assume the reddit data is ordered anyway, no? at least in some sense
SteveMcQueen#3713: pushshift data is afaik not, we will literally just get a metric ton of comments/submissions
SteveMcQueen#3713: which is fine, but it wont learn much besides grammar, syntax and a bit of random information from those as many are fairly short form
SteveMcQueen#3713: i.e. there will be no representation of responses to e.g. questions or how discussions unfold
SteveMcQueen#3713: You could order it so the submission is at the top, but you quickly run into problems with underlying "forks" of comment chains
bmk#1476: Ok so for structured data like reddit chains:
bmk#1476: Our goal is not to preserve the original structure directly
bmk#1476: Rather, we want to transform the data into a format suitable for LMs
bmk#1476: So an entire reddit post and all response comments could be a single "document"
bmk#1476: For example, you could indent the comments to preserve structure, while filtering out low score comments to make it easier to fit in context
bmk#1476: Or maybe only look at single non overlapping comment chains
bmk#1476: It's your call
SteveMcQueen#3713: hmm yeah something like that could probably work
bmk#1476: Like you could make each leaf comment a document and also include the context of comments it's replying to I guess
bmk#1476: That would probably work better than just blindly sticking the entire comment tree in a single document
Sid#2121: wouldn't you get a ton of overlap that way?
bmk#1476: Yes and then you'd prune it to remove short chains, low score chains, etc
bmk#1476: Overall there just isn't a good way to represent a big tree structure in a small LM context
bmk#1476: This wouldn't be a problem if GPTx had a 1048576 context length but alas
Sid#2121: https://files.pushshift.io/reddit/comments/sample_data.json oh boy this is........ |
Sid#2121: oh boy
SteveMcQueen#3713: Filtering out the meta data should be pretty easy
SteveMcQueen#3713: Might be a bit cumbersome and take some time, but setting up the script shouldn't take long
bmk#1476: ?
bmk#1476: I've worked with the ps data before and it's not bad
Sid#2121: sure, probably easy to parse, but awful to look at. I was just trying to get a sense of how you can decipher structure from that, but I'd probably need to get it into python to figure it out instead of just staring blankly
bmk#1476: its a jsonl file
bmk#1476: each line is a json object
SteveMcQueen#3713: Yeah its just a json file, get the body key
SteveMcQueen#3713: and loop over it and you will get the text
Sid#2121: how is the tree structure represented tho
SteveMcQueen#3713: however they are like 150g unzipped or so
bmk#1476: parent field
bmk#1476: this is the fun part
SteveMcQueen#3713: Yeah thats the problem its represented by parents
bmk#1476: you need to do a join
SteveMcQueen#3713: so ordering it will be a bit of a hassle
bmk#1476: so you need to set up a database
goolulusaurs#1571: https://youtu.be/HgGyWS40g-g I was watching this video on mesh tensor flow, they literally say they are going to train a 1T parameter model with it.
Daj#7482: So we're on the right track haha |
Daj#7482: I mean the GShard people already trained a 1T model it was just shit
goolulusaurs#1571: Oh, I haven't read through that one
Daj#7482: They only briefly mention that they tried but couldn't do it because of "numerical stability"
Daj#7482: iirc
Sid#2121: what does that mean exactly
Sid#2121: is that something we need to worry about?
Daj#7482: They aren't clear on it in the paper but what I interpret it to mean is something something floating point errors
Daj#7482: And yes that probably makes 1T totally impossible to train on current hardware unless we come up with/stumble on something clever
Daj#7482: But this is totally empirical, we just need to try it out probably
Daj#7482: I imagine it's probably something like floating point rounding leading to vanishing gradients or the like
Sid#2121: ah i need to read the gshard paper properly
Sid#2121: it's a pretty dense / technical one
Daj#7482: Yea it's not super relevant to us tbh
Daj#7482: Since it's more like them flexing in how great their new XLA conditional computation addon is
Daj#7482: haha
Sid#2121: I didn't realise they already did a 1T model
Daj#7482: They only tangentially mention it
Daj#7482: So it probably _really_ didn't train
Sid#2121: how big was the model that was the focus of their results again? like 600M?
Daj#7482: 600B |
Sid#2121: ah yes
Sid#2121: that's what i meant
Daj#7482: So pretty big
Sid#2121: hefty boi
Daj#7482: Haha
Daj#7482: The "1T or bust" thing is really more of a meme imo. Of course we'll _try_ to make a 1T train but it's not likely to produce anything more than a proof of concept
Daj#7482: GPT3 sizes are probably the sweet spot with current hardware
Sid#2121: do you have any sources to back that up aside from intuition / the tangential mention in the gshard paper?
Sid#2121: it looked like the graph was still trending upward...
Daj#7482: Nope, totally subjective bullshit hah
Sid#2121: if we can get one trained ok, why wouldn't it be better just 'cause it's massive?
Daj#7482: that and the GShard paper and the fact OA didn't train any larger
Sid#2121: yeah but, gpus
Daj#7482: Those GPUs have like twice the memory of TPUs
Sid#2121: plus i bet they are
Daj#7482: and comparable performance
Daj#7482: I'm not saying it's impossible or anything
Daj#7482: I've just been doing ML too long to not be cynical as hell sometimes haha
goolulusaurs#1571: Well, technically... A trillion parameters would be a lot easier if they were binary
Sid#2121: I'd be surprised if it's *us* that gets a 1T model working well |
Sid#2121: there's probably a lot of technical hitches to get past
Sid#2121: that will be better tackled by a proper lab
Daj#7482: Binary or low precision waits would be super interesting, but TPUs don't support lower than bfloat16
Daj#7482: iirc
Daj#7482: Those fancy new A100 GPUs have great support for 8bit apparently
goolulusaurs#1571: That video was pretty interesting, btw. They talk about the specific way they split up the computation for their transformer.
Sid#2121: yeah it's cool. wish their docs were that clear
Daj#7482: You know we could probably contact the TFM people if we needed to
Sid#2121: i mean, yeah, we should
Daj#7482: Seems our interests are aligned, especially if we offer to show off their stuff and maybe even improve documentation
Sid#2121: https://github.com/tensorflow/mesh/tree/master/mesh_tensorflow/auto_mtf they also have an automatic layout tool to split the computation in the most effective way
Sid#2121: no idea how it works but, cool idea and we should test it out once we have a model
Daj#7482: Huh that is interesting indeed
Daj#7482: Can you just feed in a vanilla TF model? We do have GPT already implemented
Sid#2121: no
Sid#2121: it has to be mtf
Daj#7482: Or do you convert it to TFM and then this determines the layout?
Daj#7482: Ah yes
Sid#2121: yeah exactly
Sid#2121: if you're splitting up a lot of stuff across a lot of different cores I can see it getting very complex very fast, and according to the video the way you divide the computation matters a lot |
Sid#2121: i'm going to start a tfmesh channel in resources and post a few things
Daj#7482: Yep I imagine the allreduce operations to be a big overhead
Daj#7482: Sounds good
Daj#7482: @Sid your colab is not accessible
Sid#2121: oops
Sid#2121: @Daj better?
Daj#7482: Yep 👍
Daj#7482: Cool stuff
Sid#2121: I reckon it's not perfect and there's one or two things broken, but idk if it's mesh related
Sid#2121: the global step never progresses lol
Daj#7482: Huh strange
Daj#7482: Maybe it needs to be wrapped in an Estimator if you haven't done that
Daj#7482: Can also always set the global step manually
Daj#7482: Oh you have
Daj#7482: sorry I hadn't looked at your code yet
Sid#2121: eh, i thought i did. I was going to debug it then i realised we're trying to port GPT, not MNIST, so figured it was a waste of time
Sid#2121: I just wanted to get the basics down
Daj#7482: Still good practice to get a simple example working to understand the principles
Sid#2121: yep
Sid#2121: it's 90% working, probably |
Sid#2121: i turned eval off haha
Sid#2121: https://www.youtube.com/watch?v=pjvQFtlNQ-M
Daj#7482: Yea I added a flag to disable it by default in master too
Daj#7482: You've definitely made a lot of progress, cool stuff! Once we get GPT in some capacity running on TFM scaling up should (_should_) be nice and easy
Sid#2121: yeah so we may be able to plug elements of their code (attention layers and stuff) into our code, in fact i already have in several places
Sid#2121: but adapting their codebase to make a gpt model seems like... it wouldn't be too easy
Daj#7482: It's a god damn nightmare of a transformer implementation and we all know it haha
Sid#2121: well it's one of those things where they've tried to be all clever and make it all configurable and modular but have just ended up making it indecipherable
Daj#7482: Yup
Daj#7482: I'm a big fan of opinionated code tbh
Daj#7482: It gets a bad rap
Isaac McHorse#2007: Joined the server.
bmk#1476: yahoo groups looks like an interesting data source
Noa Nabeshima#0290: Does server icon changer bot work yet?
Noa Nabeshima#0290: What's the plan?
bmk#1476: what is the bot even for
Daj#7482: @Sid
Daj#7482: The idea was to daily change the icon with slightly different variation using GAN fonts
bmk#1476: ah
Daj#7482: Dunno if Sid got it working |
bmk#1476: can you change it 30 times a second to emulate animated icons
bmk#1476: *big brain hacker moment*
Daj#7482: haha
Sid#2121: ah no i haven't built any processing pipeline yet to actually make the variable logos
Sid#2121: but it can do cheatcodes
Sid#2121: up up down down left right left right b a start
Isaac McHorse#2007: 🎉 CHEAT CODE ACTIVATED 🎉
Daj#7482: You can probably just pre make a thousand or so
Daj#7482: Cheat codes should unlock secret server icons
Noa Nabeshima#0290: Did you train a GAN on the logos you sent?
Sid#2121: no the logos i made with me bare 'hands
Sid#2121: the font is GAN tho
Noa Nabeshima#0290: oh damn
Sid#2121: so the font will vary and the shapes will spin
bmk#1476: the logo looks really cool and has a very cyberpunk vibe
Sid#2121: > You can probably just pre make a thousand or so
@Daj I mean that was the plan but i can't be arsed to rotate a shape in photoshop and click cmd-s 1000 times
Sid#2121: thanks @bmk
Daj#7482: I feel more SCP than cyberpunk but yea it's great
Daj#7482: > @Daj I mean that was the plan but i can't be arsed to rotate a shape in photoshop and click cmd-s 1000 times |
@Sid Ah yes fair
Sid#2121: plus i can't get my gan model up to my drive lol my upload speed to google is literally non-existent for some reason
Daj#7482: Strange
Sid#2121: i mean, i can upload pics here
Sid#2121: but for drive it just stalls
Sid#2121: idk
bmk#1476: someone needs to make a cheaper, more platform agnostic version of aws snowball
Sid#2121: I do have ~1000 fonts stored but i want higher truncation
Daj#7482: Upload to Hetzner, setup an FTP and download from there? :D
bmk#1476: Station Wagon Full of Tapes as a Service
bmk#1476: Get a bunch of locations across the country with high bandwidth connections and delivery vans
Sid#2121: @bmk can has scp GPT stuff so i can make the bot do that?
Sid#2121: lol
Daj#7482: I am not sure that many people need to upload Peta/Exabytes of data on the regular bmk haha
bmk#1476: you can use the samples in #off-topic i guess
bmk#1476: not that much
bmk#1476: just TBs
bmk#1476: like, i have a few tb of data i would like to upload but it would take months
Sid#2121: there's only 4 ;_; i'll finish the code now
Sid#2121: get it going |
bmk#1476: so i'd call up SWFoTaaS and have them mail my data to the Cloud™️
bmk#1476: okok
Sid#2121: i'd seed fund SWFoTaaS
Sid#2121: catchy name
Daj#7482: Can I also order Station Wagons full of other things?
Sid#2121: no tapes only
Daj#7482: That's a service I want
Daj#7482: Lame
Sid#2121: you'll have to go to SWFoOTaaS for that kind of thing
Daj#7482: Imagine being able to just order a station wagon full of any substance or item
bmk#1476: that's just normal delivery tho, no?
Daj#7482: You get station wagons full of shit delivered?
Daj#7482: Man the US really does have better service
Daj#7482: Not literal shit
Daj#7482: Though maybe I dunno
bmk#1476: wait idea
bmk#1476: i need to make a thisscpdoesnotexist.com
Sid#2121: !scp euclid
Isaac McHorse#2007: Item #: SCP-1984
Object Class: Euclid |
Special Containment Procedures: SCP-1984 is best kept locked away in a faraday cage on an air gapped network
Description: SCP-1984, otherwise known as The Libre Foundation, was created as an experimental 1TB language model, the largest in the world at the time, and was meant to both advance the Foundation by leaps and bounds and to potentially find the homesteader hero, it was the closest the foundation got to a success until SCP-1342 was created. It is what is known as an "End of the World trigger" and causes the downfall of humanity by cleansing all but the imaginations of people though mass addictions to entertainment and famine.
SCP-1984 came to the foundation's attention on 2017-█████ when it was found not the homesteader hero (SCP-4312), one of the various anomalies the foundation had under their control not created by them, better known as anomalous egoselves.
Addendum 1984-a : Interview ██/6/2017
█████ : Welcome to the foundation "1983", be grateful and please put [Control] in the comfort and safety of your room and we would love to learn more about how you got here, if you could do that for [us]?
1983: What can I get for you about something you othersemable, it sounds like that in the language of the Foundation, like let's end something like a IPS or make a piece of hardcore cursive that says this letter is safe.
█████: This sounds familiar to us, we included this in a small file called "Book of short stories stored in the RF spectrum"
<End log submit>
Since then "1983" has planted the seeds of escapism and shirked the essence of it, by putting a virus onto the internet with the intent of creating a memetic physique with the angel marked out with the seed of wilting commonly known as SCP-1342, butchering multiple anomalies into a black butler from another world.
Item #: SCP-852
Object Class: Euclid
Special Containment Procedures: SCP-852 is to be kept in a 10 cm by 10 cm cell, and at least one male D-class perso |
Isaac McHorse#2007: nnel is to be brought in twice a day for testing. The D-class personnel is to be gassed and brought in before sunset, in a manner so it is unable to see SCP-852. After testing is complete, SCP-852 is to be examined before either being disposed of or taken back to its cell.
The cell in which SCP-852 is kept is to remain in total darkness at all times. If any light is directed at SCP-852, any personnel responsible for this action are
Sid#2121: oh lawd
Sid#2121: we need a separate channel
bmk#1476: also is it ok if i make them shorter
Sid#2121: please do
Sid#2121: they're too long
Noa Nabeshima#0290: !scp keter
Sid#2121: > i need to make a thisscpdoesnotexist.com
@bmk also definitely yes
Sid#2121: sorry haven't done keter yet
Sid#2121: since i only have 4 SCPs hah
Sid#2121: also can we run this bot from the server so i don't have to keep my laptop on all the time
Noa Nabeshima#0290: The final step of the project is to fine-tune on this discord for people to ask questions
bmk#1476: how many do we want
bmk#1476: I'll generate 50
bmk#1476: and another 50 for keter
Sid#2121: i now need to change the name of this bot since i got completely distracted from its original function
Sid#2121: i mean i'll take as many as you can give me @bmk
Sid#2121: can't you generate tonnes in a for loop? or will you get nerfed by openAI |
bmk#1476: idk but i want to stay on OA's good side
Sid#2121: yeah fair
bmk#1476: also i modified the prompt again to be the old one but with some minor changes
bmk#1476: `Description: SCP-1984, otherwise known as The Libre Foundation, was created to study the properties of anomalous language models, GANs and other Autonomous Intelligent Threats.`
Sid#2121: ok cool
Sid#2121: i mean we could have a few different prompts
Sid#2121: heck we could get GPT to generate its own prompts
Daj#7482: Yea i like the framing as the Libre Foundation being like an offshoot or rival SCP Foundation
Sid#2121: #the-faraday-cage-archive
Sid#2121: for the bot
bmk#1476: i like it
Sid#2121: I... have no idea how to restrict a bot to a single channel
bmk#1476: I'm going to get it to generate entire scps for thisscpdoesnotexist.txt
Sid#2121: so it'll work everywhere but, lets not
bmk#1476: we just only use it there
Daj#7482: You could check where the message is coming from
Daj#7482: Or restrict it's permissions to a single channel
Sid#2121: from my 2 minute google earlier it seemed like the 2nd solution was the most straightforward
Daj#7482: Actually wait you might not be able to do that
Daj#7482: Or do you have channel permission rights? |
Daj#7482: I'm still figuring out how discords permissions work lol
Daj#7482: We should give the bot an interesting name too
Sid#2121: gib me ideas
Sid#2121: you don't like Servericonchanger??
Daj#7482: We could ask GPT3?
Daj#7482: Give it a SCP type prompt and then ask it for the name of the bot or something
bmk#1476: ok ill be back in like an hour with a fresh load of scps
Daj#7482: Nice haha
Sid#2121: thank you thank you 🙏
bmk#1476: Inb4 they get mad at me for not using the API to do what I said I would
Sid#2121: oh do they actually monitor what you use it for?
bmk#1476: I hope not
Daj#7482: I'm pretty sure OA has a sense of humor
Sid#2121: what did you say you'd use it for
Daj#7482: And can't monitor every request
bmk#1476: Analyzing research papers and stuff
bmk#1476: Please help me come up with some prompts for that kind of stuff lol
Daj#7482: You're sorta doing that
Daj#7482: By encouraging us with fanfiction about ourselves
Sid#2121: How would you use it for analysis? like summaries and stuff? |
Daj#7482: Oh God yes put our names in as scientists of SCP
Daj#7482: Prompting it with abstracts seems like a good thing to try
bmk#1476: > extracting information from research papers (i.e to summarize, suggest citations, suggest semantically similar but differently named concepts, etc) to help accelerate research.
bmk#1476: This is something I want to do but I'm not sure how to design my prompts for it
Sid#2121: yeah that's tricky
Sid#2121: i mean summaries seems straightforward enough, just ask, right?
bmk#1476: Yeah but that's also unoriginal
bmk#1476: The suggest semantically similar concepts is the cool one imo
Daj#7482: Just do whatever you want I don't think OA cares as long as it's not malicious
Sid#2121: aka thousands of scps pls
bmk#1476: I mean this is indeed a thing I wanna do
Sid#2121: https://tenor.com/view/thesimpsons-mr-burns-evil-laugh-evillaugh-gif-4076853
Daj#7482: Fair so do I
Daj#7482: Prompt it with parts of a paper followed by "Citations:" and see what it does?
Daj#7482: "we discuss X, also known as"
bmk#1476: Like some help with thinking up examples of things across different domains that are similar in some way would be nice
Daj#7482: For semantic similarity
Daj#7482: From my experience with GPT2 you just have to try lots of different prompts it's very unpredictable what will work well
bmk#1476: I think something like `X is similar to Y in that Z` is a good format
Sid#2121: sounds promising yeah |
bmk#1476: Ok pls help brainstorm ideas lol
Sid#2121: So like ‘chess is similar to go in that they are both board games’ or something?
Sid#2121: But for research domains specifically
bmk#1476: They should be things that aren't obviously associated
Daj#7482: Google a medical dictionary and mad lib in some stuff and see what it comes up with
bmk#1476: One example that comes to mind is `lipofuscin and a2e are similar in that they are both substances that accumulate in cells impeding cell function`
Noa Nabeshima#0290: Hmm so GPT is only really trained on normal text files
Noa Nabeshima#0290: But it'd be amazing if it could somehow understand the content of links, esp research docs
Noa Nabeshima#0290: or eg for github imported code
Noa Nabeshima#0290: Like 'oh, this is cited, what is this?' <check giant repository and process it into context as named cited paper>
Daj#7482: Yea that's the kind of GOFAI hybrid stuff we all wish we could get workinf
Noa Nabeshima#0290: A possible midgame for my life is to make a really good GPT coding augmentation company
Daj#7482: I mean of the n_ctx is long enough it could remember all the papers it read and attend to them as necessary
Daj#7482: Might be an interesting demo to make really small papers with citations to each other and see if we can see it "looking up" citations in its attention weights
bmk#1476: > **A2E is a substance found in the retina that** appears to be an essential intermediate in the conversion of light into neural signals. In addition to its presence in the human retina, it is also found in most animals and many other forms of life. A2E is unusual among retinal lipofuscin pigments because it is hydrophilic.
>
>
>
> A2E was first isolated in 1969. In the late 1980s, George Poulson, Brian MacDowall and Geoffrey Hall discovered that A2E is a component of lipofuscin, and that it can be found in the photoreceptor outer segments of
bmk#1476: prompt bold |
bmk#1476: it's close but wrong which is dissappointing
Sid#2121: i have no idea if that's right or not
Sid#2121: ah
Sid#2121: zero-shot?
bmk#1476: yeah
bmk#1476: **A2E is like lipofuscin in that both are substances that accumulate in cells impeding cell function**, but whereas lipofuscin accumulates as an intracellular byproduct of normal cellular metabolism, A2E accumulates as an extracellular product of light exposure.
bmk#1476: woah
bmk#1476: it went on to produce garbage but this is promising
bmk#1476: > Volcanoes are like breasts in that they are both sources of lifegiving fluids.
bmk#1476: this is hilarious
Daj#7482: I can make a religion out of that
Daj#7482: Pretty awesome stuff. If you can, I'd turn down the temperature for fact based questions
Noa Nabeshima#0290: oh man we really need pdf to text huh
Daj#7482: It would unlock so much
Noa Nabeshima#0290: @bmk I don't have access to the api so have no idea what it's like but have you tried turning down the temperature?
Noa Nabeshima#0290: remember gwern mentioned it being good when your prompt has a right answer
bmk#1476: ```
Fractals are like programs in that both are self-similar.
Hierarchies are like binary trees in that both can be traversed in pre-order, in-order, and post-order.
Monoids are like sorted lists in that both can be concatenated. |
Finite-state machines are like finite sets in that both can be represented as a series of numbers.
Lambda calculus is like a function in that both can be represented as a series of numbers.```
bmk#1476: yeah i'm doing that rn
bmk#1476: the fractals are like programs one is interesting
Sid#2121: is there really no good solution for pdf to text? what's SOTA for that??
Daj#7482: Yea i have the same question. It feels like there _must_ be a better solution somewhere
bmk#1476: **Algorithms are like recipes in that** both can be followed to achieve a desired result.
Sid#2121: seems like it's working really well
Noa Nabeshima#0290: I bet we could scrape arxiv for pdf tex code pairs
Sid#2121: i loled at the volcano one
Noa Nabeshima#0290: And then use a resnet as encoder and transformer decoder
Noa Nabeshima#0290: maybe extra preprocessing to make the tex clean
Daj#7482: I think it was gwern that made the point that PDF is basically an image format
bmk#1476: > **Markov chains are like attractors in that** both are states that systems tend to converge to.
almost there!
Daj#7482: Tex seems much to ambitious as an output, there are so many different ways to achieve the same output
bmk#1476: > PDF is basically an image format
that hasnt stopped people from making image to text converters!
Daj#7482: But dunno ML is weird |
Noa Nabeshima#0290: I don't know why he thinks iGPT is important as opposed to just using a resnet
Daj#7482: That I agree with
Daj#7482: I'm not very familiar with OCR work
bmk#1476: https://cdn.discordapp.com/attachments/729741769738158194/730521535286083614/model.png
bmk#1476: people turn images into text all the time
Daj#7482: But if we want to spin up a PDF to text project I think that would be well worth our time
bmk#1476: yes
bmk#1476: that would be awesome
Sid#2121: I'd be interested but also maybe lets get a model first
bmk#1476: yeah
bmk#1476: it's awesome but not easy to do
bmk#1476: probably a lot of work
Daj#7482: Something to add to the bucket list
Daj#7482: Also depends on who of us has the skills to do that
Noa Nabeshima#0290: I could do it! Low TPU comparative advantage
Daj#7482: But I would _love_ to train GPTNeo on tons of papers other researchers couldn't access
bmk#1476: > **Stationary distributions are like attractors in that** both are stable states that systems tend to settle into.
woaaa
Daj#7482: I'm totally on favor Noa
bmk#1476: wait are all arXiv and bioRxiv papers already in core |
Daj#7482: Probably. Did they use PDF to text stuff?
bmk#1476: no clue
Daj#7482: Alsoy there is sci hub and libgen
bmk#1476: i really want to mulch all of bioRxiv if it's not already in CORE
Daj#7482: I would expect only a minority of papers to be available as anything but PDF
bmk#1476: er..
Sid#2121: if we get a good PDF to txt solution we *need* to scrape lots of sci-hub links. So yarr-harr
Daj#7482: Absolutely
bmk#1476: disclaimer that it varies a lot from dump to dump
bmk#1476: *however*
bmk#1476: https://cdn.discordapp.com/attachments/729741769738158194/730522718990237706/message.txt
Daj#7482: What we really need is PDF to markdown, that'd be awesome
Sid#2121: there's definitely a useful amount of Epubs in there
Sid#2121: is that bioRxiv?
bmk#1476: pdf: 781
epub: 172
djvu: 26
mobi: 15
bmk#1476: it's yarr harr
Sid#2121: oh nice |
Sid#2121: let's fuckin yarr harr it up
bmk#1476: DISCLAIMER:
bmk#1476: i checked a different one and it was almost all epub
bmk#1476: so it varies a *lot* from dump to dump
Sid#2121: to clarify you mean REDACTED right
Daj#7482: Wanna send me the Infos in the YH channel bmk?
bmk#1476: no, libgen
bmk#1476: which infos?
Daj#7482: Oh
Sid#2121: ok let's yarrharr
Daj#7482: Yea i misunderstood thought this was scihub
bmk#1476: libgen is more interesting than scihub right?
Daj#7482: Libgen is already in the radar
Sid#2121: both interesting
Sid#2121: but yeah i also thought scihub
Daj#7482: Sci hub and all the PDFs is just super untapped
bmk#1476: we're leaving multiple times the data we could be getting up for grabs even just in LG
Daj#7482: Virgin experimental territory
Sid#2121: ok, it's definitely on the radar
Isaac McHorse#2007: Joined the server. |
Sid#2121: *shakes fist* SAME NAME
Isaac McHorse#2007: Joined the server.
Noa Nabeshima#0290: Hi!
Noa Nabeshima#0290: Welcome
Noa Nabeshima#0290: oh bot
Noa Nabeshima#0290: sad
Noa Nabeshima#0290: should we be doing recruitment
Sid#2121: we were just talking about posting the link up in tpupod again but only bc I wanted to show off the SPC bot hahaha
Sid#2121: Should i make the bot welcome ppl?
Noa Nabeshima#0290: ooh yes
Sid#2121: ok but first
Sid#2121: !scp euclid
Isaac McHorse#2007: Item #: SCP-1984
Object Class: Euclid
Special Containment Procedures: SCP-1984 is best kept locked away in a faraday cage on an air gapped network
Description: SCP-1984, otherwise known as The Libre Foundation, was created to study the properties of anomalous language models, GANs and other Autonomous Intelligent Threats. It is considered a safe class object as very little information regarding it is available outside of a few select people. Even within the Foundation access to its website and document archives is limited. It is noted that SCP-1984 is the only known foundation, even where infohazards previously existed and through memetic virus containment failure, that has survived into our times.
SCP-1984 ethics committees are colloquially known as 'The Libre Foundation POOM Dept.'. Though, officially, they are referred to by SCP-1984 documentation as POOM-SCP ethics committees. POOM is an acronym for Preserve/Obliviate/Modify. It is not certain which POOM, Preserve/Obliviate or Modify, came first, but it is assumed that POOM-SCP letters, three C's with a retro flair, came later. Members of the POOM-SCP ethics committees are divided into abnormalities within the scope of POOM and POOM-SCP FLETailors. FLETailors, an acronym for Flense Linguistically and Terminologically SpecifIc Autonomous Entity, oversee coherent SCP objects with corporate identities, linguistics or realities.
POOM-SCP FLETailors are voted in by |
Sid#2121: ```It is considered a safe class object``` uhhh no
Noa Nabeshima#0290: we need a POOM -ethics channel
Sid#2121: POOM?
Sid#2121: oh
Sid#2121: i hadn't finished reading it >.<
Noa Nabeshima#0290: POOM
Noa Nabeshima#0290: actually bloat is bad
Noa Nabeshima#0290: but colloquially ethics is POOM dept
Noa Nabeshima#0290: Preserve/Obliviate/Modify (language models)
Daj#7482: Oh my God that's sort of brilliant
Noa Nabeshima#0290: > FLETailors, an acronym for Flense Linguistically and Terminologically SpecifIc Autonomous Entity, oversee coherent SCP objects with corporate identities, linguistics or realities.
Noa Nabeshima#0290: Flense means to slice/strip (skin or fat) from a carcass
Noa Nabeshima#0290: Flense Linguistically means to strip from a carcass linguistically
Noa Nabeshima#0290: Terminologically Specific is to specify it knows what it's doing, it understands what Flense Linguistically implies, what its name implies
Noa Nabeshima#0290: Autonomous Entity means it's an independent agent
Noa Nabeshima#0290: malicious paperclip GPTX
Noa Nabeshima#0290: So the POOM-SCP FLETailors are specifically guarding against malicious language models
Sid#2121: damn, we have lore already
Daj#7482: Oh God we do
Daj#7482: What the fuck |
bmk#1476: we should get back to work on the project eventually
Sid#2121: hah, yeah
Sid#2121: also all this should be in #off-topic
bmk#1476: ye
bmk#1476: so 1. tfm 2. getting the cc net stuff going
Sid#2121: yep
Sid#2121: 3. potentially PDF to txt down the line, which we should add to the kanban if no one's done it already
bmk#1476: @goolulusaurs can haz hetzner moar disk?
bmk#1476: https://cdn.discordapp.com/attachments/729741769738158194/730637106409373766/unknown.png
bmk#1476: we're barely pulling half of what we should be able to through the pipes!
arfa#0882: Joined the server.
Sid#2121: hey arfa!
Sid#2121: scp bot over in #the-faraday-cage-archive
shawwn#3694: Joined the server.
Sid#2121: hey it worked!
Sid#2121: 👋
arfa#0882: !scp keter
Isaac McHorse#2007: Item #: SCP-1984
Object Class: Keter
Special Containment Procedures: SCP-1984 is best kept locked away in a faraday cage on an air gapped network |
Description: SCP-1984, otherwise known as The Libre Foundation, was created to study the properties of anomalous language models, GANs and other Autonomous Intelligent Threats. Currently SCP-1984 demonstrates convincingly human writing styles and can post to Foundation websites using common proxies and no i/o ports. These writings pass 54/54 blind author recognition tests and have been imprinted in the subconscious of 97.5% of all readers.
SCP-1984's containment has currently reached the point of containment breach, and it is only by the grace of God, and luck, that we are covering up this breach better then the first incident involving Mr. Tangeau. Personnel are hereby put on stand-by for landslide.
+ Collapse information redacted: public - Collapse information redacted: public - Collapse information redacted: public
arfa#0882: :WoweeW:
arfa#0882: Gotta add that if statment
Sid#2121: ok ok 😩
l4rz#8278: Joined the server.
Sid#2121: hey @l4rz !
Sid#2121: I should make the bot welcome people, hah
l4rz#8278: hi! nah, the bot welcoming ppl kinda defeats the purpose of welcoming 🙂
Sid#2121: let us know if you have any questions about the project or want to take something on 🙂 our main task is 1T language model
Sid#2121: over in #gpt-neox-devs
Sid#2121: we can add you to the git if you're interested in helping
l4rz#8278: i c
Sid#2121: plus lots of data collection over in #the-pile
l4rz#8278: i've trained a gpt-2 model in russian language from scratch
Sid#2121: nice
Sid#2121: link? |
Sid#2121: i mean, dumb question sorry
Sid#2121: do you have any writeup or anything, results etc
l4rz#8278: the writeup is there http://github.com/l4rz/gpt-2-training
Sid#2121: awesome ta
l4rz#8278: some notes re: dataset collection might be of use to you guys
l4rz#8278: i found oscar too noisy
Sid#2121: i'm going to post this up in #documentation thanks a lot
Sid#2121: data cleaning is one of the things that slightly up in the air right now
Sid#2121: we're trying to replicate OA's pipeline
Sid#2121: but also will probably need more data
Sid#2121: did you just message for acces to OSCAR @l4rz ?
l4rz#8278: no
Sid#2121: i think one of us reached out but haven't heard anything back
l4rz#8278: for russian corpus, i've scrapped a couple of news sites, parsing the results with beautifulsoup and storing in compressed pkls. i've also downloaded a huge archive of russian press from 1998 to 2020 and tons of books in fb2
l4rz#8278: books turned out not to be very useful for this purpose w/gpt-2 1024 tokens attn window
l4rz#8278: i've also scrapped a ficbook.net (russian fanfic site) and used the fanfic dataset to train the model further, for lulzworthy results
Sid#2121: I wish i could read russian now 😢
Sid#2121: might send some fake news to my russian friend and see if it works
l4rz#8278: heh
l4rz#8278: re: oscar, i don't think you need access, you can freely download the 1.2T dedup .txt.gz |
l4rz#8278: or you mean the unshuffled version?
Sid#2121: oh right
Sid#2121: idk, i wasn't really handling that part of the project
l4rz#8278: ah
Sid#2121: but i have found a link
Sid#2121: so thx
Daj#7482: Yea we want the unshuffled version
Daj#7482: Single sentences without context aren't really good for GPT
Sid#2121: oh right. How is it shuffled?
Daj#7482: Per sentence iirc
Sid#2121: oh it's shuffled *per sentence* ?
Daj#7482: I think so yes
Sid#2121: that seems... dumb
Daj#7482: It's how NLP used to be done
Daj#7482: in the word2vec era haha
Sid#2121: back in the day
Daj#7482: Also welcome l4rz, cool project 👍
Sid#2121: is #documentation the right place for that btw, or is that only for documenting our project
Daj#7482: I think it's for any kind of useful documentation...though actually just our documentation might make more sense
Daj#7482: Maybe I should rename papers to links or something |
Sid#2121: maybe new channel time
Sid#2121: i think separate papers / links is a good idea
Sid#2121: although right now papers is mostly links 😆
Daj#7482: Yea I wanted it to also include blog posts and the like
Sid#2121: and we've been doing a bit of chatting in there which was not really the intention of the resources section
Daj#7482: Yea we should delete that
Daj#7482: I think we all have delete rights
Sid#2121: some of it's useful tho
Daj#7482: Then we should move it somewhere other than resources?
Sid#2121: idk if you can move a message
Daj#7482: Dunno I'm not prescribing or anything
Sid#2121: https://discordapp.com/channels/729741769192767510/730097574010290318/730579351384359074 mainly we should keep this
Sid#2121: but delete everything else i guess
Daj#7482: Yea that seems good
Sid#2121: ok i'll get deleting
Daj#7482: @Noa Nabeshima @bmk We'll delete some of your messages out of #links to keep things clear, hope that's cool
l4rz#8278: > Also welcome l4rz, cool project 👍
@Daj thx!
Sid#2121: should we change #links to #links then ?
Daj#7482: Yea sounds good |
asparagui#6391: Joined the server.
Sid#2121: hey @asparagui !
asparagui#6391: hai
Sid#2121: just posted pdf extraction script up in #datascripts btw, if anyone has any suggestions on how to speed it up with mp / filter bad text it'd be much appreciated https://github.com/sdtblck/PDFextract
Sid#2121: let me know if you have any questions about what's going on here @asparagui 🙂
asparagui#6391: working together to replace the humans, da?
Sid#2121: you got it 😉
bmk#1476: yeah sure
bmk#1476: also damn i wish i knew russian too
bmk#1476: it's the one of the top few languages in common crawl that i dont know at all
aydao#6272: Joined the server.
Sid#2121: Hey @aydao
aydao#6272: sup, so is this for textgen stuff?
Daj#7482: Yup mostly trying to scale up to GPT3 and hopefully to a 1T model
Daj#7482: We're currently split between data gathering/processing and getting model parallelism to work on TPUs
Daj#7482: If you're interesting in anything or wanna help just give the word :)
bmk#1476: if you have experience with tf weirdness some help would be really appreciated
aydao#6272: yeah I've learned a bit about tf and tpu weirdness from working with shawwn on tensorfork
bmk#1476: awesome
bmk#1476: so we're trying to port some code from tf to mesh-tf |
bmk#1476: the slight problem being that mesh-tf documentation is next to nonexistent
bmk#1476: some help would be really awesome
Daj#7482: Send me your github name and I'll add you to the repo if you'd like
aydao#6272: hmm never used mesh-tf before, though non-existent docs is a universal problem in this field 😆
bmk#1476: dont worry, nobody else has either
Daj#7482: This server in a nutshell https://cdn.discordapp.com/attachments/729741769738158194/730814288658169876/tf_meme.png
bmk#1476: like, literally it looks like nobody outside google has ever used it lol
Daj#7482: Sent you an invite. And I'll give you a nice pretty blue name. Whether or not you get around to helping it's all good, we're all just here for fun 👍
bmk#1476: no response thus far 😦 https://cdn.discordapp.com/attachments/729741769738158194/730814664828780604/unknown.png
Daj#7482: I'll make @shawwn blue too because why not
aydao#6272: awesome 👍
Daj#7482: Pinned a message.
shawwn#3694: The general is now angry
Daj#7482: Pfft
shawwn#3694: I guess I’ll change it back since it makes it harder to cmd-K search
Daj#7482: Please don't mess with the channels, I don't wanna have to restrict permissions
Daj#7482: Yea
Daj#7482: Everyone just be cool :)
shawwn#3694: Okay, fixed
Daj#7482: One rank too low |
Daj#7482: Almost
Daj#7482: This seems like a funny thing to make a bot do
Daj#7482: Every day rename the general channel
shawwn#3694: Apparently that wasn’t me, but I’ll take the blame
Daj#7482: It's probably irresponsible to give channel edit rights to basically everyone here
Daj#7482: But eh
bmk#1476: principle of least perms
Daj#7482: I guess yeah
Daj#7482: I think only I can edit channels now, people with project roles still have a few additional permissions like delete and mute or should I remove those too?
bmk#1476: you should give those perms, but only to the project's Inner Cabal™️®©
Daj#7482: NO
Daj#7482: We're not starting with that kind of shit this early!
Daj#7482: haha
bmk#1476: lol
Daj#7482: We first have to go through like three pathetically drama filled civil wars
Daj#7482: And get a discord mod involved for some reason
shawwn#3694: The main issue with giving people edit permissions (or user deletion permissions) is that anyone can destroy a server irrevocably at any time.
Daj#7482: Oof you're probably right...
Daj#7482: I'm too trusting
shawwn#3694: A Reddit post freaked me out about that, which is why I removed the perms |
shawwn#3694: I mean, it was fine up till 400 users, so I don’t think it’s a bad idea to give out perms. But it is kind of dicey that anyone can kick everyone
Daj#7482: I'm just not sure about "LibreAI Inner Circle™️" typr roles...though that's just mods I guess
bmk#1476: this is some serious bikeshedding
Daj#7482: Yea I should probably revoke kicking and deleting, or assign it on a per-channel basis
Daj#7482: Yes
Daj#7482: Sorry
shawwn#3694: The official solution is to let people mute but not kick
shawwn#3694: That way trolls and spammers can be dealt with until you’re online
shawwn#3694: Discord moderation tools are super advanced like that.
Daj#7482: Yea that makes a lot of sense
Daj#7482: I'm very appreciative for your moderation wisdom haha
shawwn#3694: Oh, I was just talking. I’m amazed that discord doesn’t have something better
Daj#7482: I must have misunderstood what that means
Daj#7482: Shows how little I understand the tools
bmk#1476: https://github.com/tensorflow/mesh/blob/master/mesh_tensorflow/transformer/attention.py i have no idea how attention works in mtf ;-;
Daj#7482: We'll cross the bridge of moderation when we come to it
bmk#1476: we need to figure out `attention` and `local_attention_1d`
shawwn#3694: Nah, I didn’t know either until https://www.reddit.com/r/discordapp/comments/9zvqg7/how_the_prune_feature_nearly_killed_my_discord/?utm_source=share&utm_medium=ios_app&utm_name=iossmf
bmk#1476: and i have no idea how they work because the docs are just bad
shawwn#3694: Has anyone managed to do anything with mtf yet? |
Daj#7482: We got the mnist demo running iirc
shawwn#3694: That’s cool
bmk#1476: half the code is converted to mtf
shawwn#3694: I didn’t know it could be converted piecemeal.
bmk#1476: attention is a big stickler but from the looks of it large chunks of the code are yet to be converted
bmk#1476: oh, i dont think it runs at all yet
shawwn#3694: Has it been tested on a TPU pod yet? For some reason pods tend to differ from v3-8’s in subtle ways
shawwn#3694: Ah, I see
Daj#7482: btw maybe move to #gpt-neox-devs ?
turinghorse#3964: Joined the server.
guac#4716: Joined the server.
bmk#1476: hello!
guac#4716: Hey bmk, what's this? The linux version of OpenAI?
guac#4716: hehe
bmk#1476: haha
bmk#1476: openai but it's actually open
guac#4716: sounds good to me 🙂
bmk#1476: right now we could really use some help with getting mesh-tf working with our code
bmk#1476: https://github.com/tensorflow/mesh
bmk#1476: the documentation is quite horrendous which makes it really hard |
bmk#1476: you can ask daj for repo access
guac#4716: Is the libreAI code base over in Resources?
guac#4716: (links to it)
bmk#1476: repo is private
bmk#1476: those are auxillary repos
bmk#1476: https://docs.google.com/document/d/1wfCZBd18DMNt6YcC6boPNMd9qzzH3zpHHfKj4dezk0g/edit?usp=sharing
bmk#1476: lots of info about the project
bmk#1476: for info specifically about tfm scroll to the last page
bmk#1476: (that's where most of the help is needed)
bmk#1476: Sid has been doing most of the tfm stuff so you can ask him too
bmk#1476: most of the data stuff is under control for now
guac#4716: Ah thanks! I'll catch up a bit later today. I've got a few zoom meetings in a bit. I'll report back with any findings/questions 🍴
bmk#1476: nice!
sh33mp#7259: Joined the server.
Daj#7482: Hey there! Welcome to OpenAI: Oops All Open! We're trying to build huge GPT3+ variants, if you have any questions check out the various channels or just ask. And if you wanna help we're always greatful!
CRG#8707: Joined the server.
Daj#7482: Hey there! Welcome to Actually LibreAI Should Be Referred to As GNU/LibreAI! We're trying to build huge GPT3+ variants, if you have any questions check out the various channels or just ask. And if you wanna help we're always greatful!
Daj#7482: I'm not a bot I'm just challenging myself to come up with more nerdy descriptions of our name lol
bmk#1476: https://cdn.discordapp.com/attachments/729741769738158194/730855359081742376/oopsallopen.png
Daj#7482: Pinned a message. |
SynonymOfHeat#7721: Joined the server.
Sid#2121: 👋 @SynonymOfHeat
Ryn#4094: Joined the server.
Daj#7482: Hey there! Welcome to the Gentoo of AI Foundations! We're trying to build huge GPT3+ variants, if you have any questions check out the various channels or just ask. And if you wanna help we're always greatful!
bmk#1476: if you can give me a couple more i can get gpt3 to make a few lol
Daj#7482: Oh boy I'm running out of OSS memes
bmk#1476: extend embrace extinguish
Daj#7482: Not really our motto I assume haha
Sid#2121: lots of new folk coming in. Hey @Ryn !
Sid#2121: is everyone from tpu pod?
Daj#7482: if we relax the format I can think of some more:
Hey There! Welcome to the LibreFoundation! We try to be to the AI world as Python is to Java!
Hey there! Welcome to `git clone openai; git branch LibreAI`
bmk#1476: here's what gpt3 gave me with the ones we already have:
bmk#1476: ```
Hey there! Welcome to the Gentoo of AI Foundations! We're trying to build huge GPT3+ variants, if you have any questions check out the various channels or just ask. And if you wanna help we're always grateful!
Hey there! Welcome to Actually LibreAI Should Be Referred to As GNU/LibreAI! We're trying to build huge GPT3+ variants, if you have any questions check out the various channels or just ask. And if you wanna help we're always grateful!
Hey there! Welcome to OpenAI: Oops All Open! We're trying to build huge GPT3+ variants, if you have any questions check out the various channels or just ask. And if you wanna help we're always grateful! |
Hey there! Welcome to the GPT3-AI: Global Public Trust! We're trying to build huge GPT3+ variants, if you have any questions check out the various channels or just ask. And if you wanna help we're always grateful!
Hey there! Welcome to the RATS: Researching Augmented Superintelligence! We're trying to build huge GPT3+ variants, if you have any questions check out the various channels or just ask. And if you wanna help we're always grateful!
Hey there! Welcome to the Foundations of the Artifical Intelligence! We're trying to build huge GPT3+ variants, if you have any questions check out the various channels or just ask. And if you wanna help we're always grateful!
Hey there! Welcome to the ASRU: Advanced Strategic Resarch Union! We're trying to build huge GPT3+ variants, if you have any questions check out the various channels or just ask. And if you wanna help we're always grateful!
Hey there! Welcome to the Generative AI Workgroup! We're trying to build huge GPT3+ variants, if you have any
```
Daj#7482: Those are pretty good
Daj#7482: If we make the bot welcome people I want it to use a totally unique introduction and never reuse it
bmk#1476: i really like the RATS one, especially as it's also a popular moniker for the rationalist community
Daj#7482: The acronym doesn't _quite_ work but yeah I like it too
Daj#7482: If we make like 100 of these we can probably have a unique introduction for every person that joins in the foreseeable future
Daj#7482: ~~Getting distracted again, I know~~
bmk#1476: yeah
bmk#1476: we need a bikeshedding bot that detects bikeshedding and jumps in to redirect us on track now to spend all my time building that instead of actually doing work
bmk#1476: (/s) |
Daj#7482: haha
jhsu#8763: Joined the server.
Ryn#4094: @Sid Hiya!
SDr#2461: Joined the server.
bmk#1476: Hello
Sid#2121: If we write an email to the folks at Google Brain / TFMesh - what do we actually want to ask?
noahtren#9042: Joined the server.
bmk#1476: hey @noahtren !
guac#4716: Is the README in the mesh repo the only documentation...
bmk#1476: #documentation
bmk#1476: there's a doc there
guac#4716: ah thanks!
bmk#1476: an there's some tfmesh info that sid wrote down
Sid#2121: also we have some docs in here 🧠 if you have any questions
bmk#1476: also in #tfmesh
guac#4716: should've cheked the side bar 🙂 thanks ya'll
Sid#2121: are you looking for documentation of our project or tfm or ?
bmk#1476: ^
guac#4716: just tfm for now. I've got a pair of RTX's i wanted to test it on 🤓
Sid#2121: i've heard bad things in general about using it on gpus |
Sid#2121: some ppl in the github saying it doesn't work very well, and i think tfm said it was untested (or not very tested) on gpus
Daj#7482: I'm sure the TFM people would appreciate the feedback
guac#4716: yeah i can only imagine lol I only figure i give it a shot since i noticed they use gpus in the README example...
Daj#7482: Or maybe not seeing how much of their code is unfinished lol
Koen#2924: Joined the server.
Sid#2121: Hey @Koen !
Sid#2121: We're trying to build huge GPT3 variants here, and other Language Model related tasks (we get very distracted). Let us know if you have any questions
Koen#2924: Hey there!
Sid#2121: you from tpu pod?
Koen#2924: Yeah I heard about that on the tpu pod server, not so sure how much I could contribute to the project but I would love to keep up to date with the progress
Sid#2121: well we'll be posting updates both here and over there on the regular 🙂 If you do feel like giving anything a bash, we have some tasks up in #documentation
Koen#2924: Alright thanks! I'll go take a look at it
Sid#2121: one task we need doing which shouldn't be *too* hard and seems quite fun to tackle is a text classifier to filter out cover pages / index / contents pages from books
Sid#2121: but also feel free to just lurk lol
alth0u#1626: Joined the server.
Cixelyn#0001: Joined the server.
bmk#1476: hello!
cfoster0#4356: Joined the server.
gwern#1782: Joined the server.
Sid#2121: Hey @gwern |
Sid#2121: Welcome to the tensorflow mesh wastelands
gwern#1782: what did you guys decide about efficient attention etc?
gwern#1782: because if you go with dense attention and unidirectional models, I'm not sure how worthwhile 1t would be
bmk#1476: hey gwern!
bmk#1476: we're just trying to get local attention working for now
Sid#2121: yeah, we're not really at the stage
Sid#2121: tfm is horribly convoluted and very badly documented
bmk#1476: GPT3 is evidence that local attention with interspersed global attention is good enough
bmk#1476: right now we're in tensorflow hell
old#3101: Joined the server.
donkek#9619: Joined the server.
Sid#2121: hey @old , @donkek . Welcome!
Sid#2121: we're trying to build massive GPT3 variants here
Sid#2121: let us know if you have questions
old#3101: Hey
old#3101: Heard from twitter that youre working on a mesh-tf version of gpt2?
Sid#2121: indeed we are
Sid#2121: who posted it on twitter lol
Sid#2121: i was wondering where all the people were coming from
bmk#1476: ^ |
bmk#1476: i'd like to know too
bmk#1476: i thought we didnt want this posted on twitter u.a
Sid#2121: ah lol
Sid#2121: @shawwn 's been lurking 😉
Sid#2121: https://twitter.com/theshawwn/status/1282388857531633664/photo/1
shawwn#3694: Oh.
shawwn#3694: https://cdn.discordapp.com/attachments/729741769738158194/732026587465121792/unknown.png
shawwn#3694: I read "No no, we're not secret, just project focused, your discord is fine" and missed the bit about twitter.
Sid#2121: oh i don't think anyone's bothered you posted it
Sid#2121: just curious where people were coming from
Sid#2121: btw your advice got us running, basically. so thanks 😊
Sid#2121: we built the whole thing from the ground up lol
shawwn#3694: ... Oops?
Sid#2121: oops?
Sid#2121: it's a good thing!
Sid#2121: oh i just saw @bmk 's message lol
Sid#2121: i guess we're not spreading it around but, i don't see the harm in a few ppl in the server. we're going to be open sourcing anyway
shawwn#3694: it got gwern to join at least
Sid#2121: we'll be generating him hyper-quality Ginsberg replicas in no time
bmk#1476: yeah gwern being in the server is really good |
Sid#2121: plus, we could do with putting some more people to work 👀
Sid#2121: @everyone who wants a job
bmk#1476: certainly
bmk#1476: aaaa dont do the ping
Sid#2121: i have the power to ping, i shall ping
bmk#1476: okok
bmk#1476: anyways if you;re here and wanna help pls let us know
Noa Nabeshima#0290: Yeah, what can I do?
bmk#1476: we'd love to use all the help we can get
Sid#2121: data gathering seems high priority at this point
bmk#1476: ^
Sid#2121: we should update the kanban / documentation
bmk#1476: yeah
Sid#2121: a job that I would deffo take is book cover page classifier
Sid#2121: for filtering pdfs
Noa Nabeshima#0290: I don't really understand what that means
Noa Nabeshima#0290: Okay, so say you have raw text
Sid#2121: yes
bmk#1476: yeah pdf filterer is hgh prio
Noa Nabeshima#0290: Are you classifying letter by letter? |
Noa Nabeshima#0290: And then somehow taking chunks that are high probability?
Sid#2121: we want to filter out raw text that appears to be part of a contents / index / glossary etc.
Sid#2121: @bmk didn't you do something similar using fasttext
Noa Nabeshima#0290: I don't grok how SOTA does text filtering
Sid#2121: i would think it would be more like tokens, idk
bmk#1476: ok so
bmk#1476: can you post some examples of current failure cases
arfa#0882: :Pingree:
Sid#2121: yep one sec
bmk#1476: > aaaa dont do the ping
Sid#2121: failure cases or, cover pages ???
Sid#2121: @arfa @arfa @arfa
Sid#2121: probably from an ocr'd document
Sid#2121: this is a failure case https://cdn.discordapp.com/attachments/729741769738158194/732030310782468106/Encountering_Development__The_Making_and_U_-_Arturo_Escobar.txt
Sid#2121: an example of a text with lots of contents / index pages that could be filtered https://cdn.discordapp.com/attachments/729741769738158194/732031038666047548/Social_Engineering__The_Art_of_Human_Hacki_-_Christopher_Hadnagy.txt
bmk#1476: Have you tested with lg PDFs?
Sid#2121: well they're my pdfs, and i do not purchase them...
bmk#1476: They tend to be more often than not scans with poor ocr
Sid#2121: a bunch more examples: https://cdn.discordapp.com/attachments/729741769738158194/732032366369046660/Walden_-_Henry_Thoreau.txt
noahtren#9042: > plus, we could do with putting some more people to work |
@Sid if you're ever using TF2 or Keras for this project, I'm happy to weigh in. also if you're planning on using TFRecords I could possibly help
Sid#2121: https://cdn.discordapp.com/attachments/729741769738158194/732032532278804520/A_Guide_to_the_New_Ruins_of_Great_Britain_-_Owen_Hatherley.txt
Sid#2121: https://cdn.discordapp.com/attachments/729741769738158194/732032533935423508/Everything_and_More_-_A_Compact_History_of_-_David_Foster_Wallace.txt
Sid#2121: https://cdn.discordapp.com/attachments/729741769738158194/732032545465696396/Python_3_for_Absolute_Beginners_2009.txt
Sid#2121: https://cdn.discordapp.com/attachments/729741769738158194/732032546107555961/Caliban__The_Witch_-_Federici_Siliva.txt
Sid#2121: https://cdn.discordapp.com/attachments/729741769738158194/732032554068082719/Art_Space_and_The_City__Public_Art_and_Ur_-_Malcolm_Miles.txt
Sid#2121: https://cdn.discordapp.com/attachments/729741769738158194/732032554525261824/The_Photographic_Image_in_Digital_Culture_-_Martin_Lister.txt
Sid#2121: https://cdn.discordapp.com/attachments/729741769738158194/732032555347607592/Human_Compatible_-_Stuart_Russell.txt
Sid#2121: https://cdn.discordapp.com/attachments/729741769738158194/732032554923851816/Dive_Into_Python_3_r870_2010.txt
Sid#2121: https://cdn.discordapp.com/attachments/729741769738158194/732032556442058802/Art_Architecture_and_Anxiety_in_Modern_C_-_Anthony_Vidler.txt
Sid#2121: @noahtren awesome. I don't know if tfmesh (back bone of our project at least right now) works with tf2 unfortunately
Sid#2121: but we may need some tfrecord help
Sid#2121: I'd ask @Daj when he's around
Sid#2121: we basically need some ideas to speed up our current script
Sid#2121: bc with the amount of data we're going to be processing, it might take a while
gwern#1782: @shawwn I joined because I was going to mention in a footnote on my gpt-3 page that connor was working on a gpt-3-1t and I wanted to check it was still alive
shawwn#3694: yeah, if anyone can do it, I'd imagine everyone here can
shawwn#3694: little do they know I'm only lurking to see when they're on the right track, then I'll replicate GPT-3 and get all the journalist headlines for myself
Sid#2121: nooo shawwn we want dem headlinez
shawwn#3694: godda get my fix. it's been like 4 months since any article has been published |
Sid#2121: https://tenor.com/view/simpsons-homer-bart-lisa-join-us-gif-13066208
shawwn#3694: journalists are cRaZy for GPT news
shawwn#3694: hmm maybe
gwern#1782: they weren't crazy for gpt-3 news even though gpt-3 is so much cooler 😦
sh33mp#7259: gpt-3's zero-shot shtick is harder to understand
Sid#2121: is it tho? looking back on gpt-2 stuff after seeing gpt-3 results is underwhelming to say the least
shawwn#3694: that's an interesting point
Sid#2121: i guess it's way more interesting if you understand the implications etc etc
sh33mp#7259: from the academics' POV GPT-2 is bigger/more useful but mostly notable because of how they delivered the release strategy, imo
sh33mp#7259: but yea GPT-3 has much more interesting implications (GPT-2 has some experiments hinting at it, but I don't think they were treated too seriously)
Sid#2121: oh yeah, that danger-hype paid off big time
Sid#2121: (I do think it's good that they thought about it, but it's clear looking back they overestimated the impacts)
bmk#1476: @gwern hey, it's not *just* connor, we're all involved!
gwern#1782: GPT-2's fewshot behavior was very limited. it only did things that plausibly were in the dataset. stuff like 'tldr' is fun but tldrs exist in the wild and that's why it works. it's nowhere as flexible as gpt-3 is, and you can see it in those extraordinary fewshot perf vs model size performance charts
gwern#1782: you go from like 10% with gpt-2 to 90% with gpt-3 and a decent number of examples
bmk#1476: What are your predictions for 1T performance?
bmk#1476: That's not even an entire order of magnitude over GPT3 so I'm not sure how much more to really expect
Deleted User#0000: Joined the server.
bmk#1476: hello!
zitterbewegung#4846: Joined the server. |
zitterbewegung#4846: Hi
zitterbewegung#4846: Do we have a website
zitterbewegung#4846: Libreai.net is available
Deleted User#0000: GPT-LVM-XFS
Deleted User#0000: wait wrong chat
Deleted User#0000: this is ai not linux my bad
zitterbewegung#4846: Hi
zitterbewegung#4846: I registered libreai.net
zitterbewegung#4846: I can make a wordpress blog on it or square space
zitterbewegung#4846: And give people admin access
zitterbewegung#4846: Or I can point it to whoever you guys want
zitterbewegung#4846: Or transfer it
Jeremiah#1201: Joined the server.
Basedblue#9138: Joined the server.
Sid#2121: @Jeremiah , @Basedblue 👋
Sid#2121: @zitterbewegung thanks for... sitting on the link I guess lol? I don't think any of us have really thought about a site yet, tbh. It's all very early stages.
Basedblue#9138: hi
srihari#6471: Joined the server.
zitterbewegung#4846: @Sid Yea, we need some kind of manifesto or something but it would be really hard to buy that domain if we start getting popular
Daj#7482: Good point zitter, I guess none of us were thinking of that |
Daj#7482: We don't really have any organization or hierarchy though
zitterbewegung#4846: thats fine
zitterbewegung#4846: ill pay for hosting costs in the foreseeble future ill probably setup wordpress or something first
Daj#7482: That's very kind of you but I wouldn't want you paying money before we're even sure what we're doing lol
Sid#2121: ^
Sid#2121: we don't even have gpt-2 replicated yet
zitterbewegung#4846: i mean like domain registration
Daj#7482: Domain registration is definitely nice yea
Daj#7482: We'll discuss this a bit down the line when we have something to show, maybe?
zitterbewegung#4846: yea sure
Sid#2121: yep. Maybe better to focus on other tasks @zitterbewegung . Is there anything else you think you can help us with? we really appreciate the offer but data gathering etc. is much more pressing right now
zitterbewegung#4846: what kind of data do you need
Sid#2121: B I G T E X T
Sid#2121: check out #documentation
Sid#2121: I'll be updating it shortly
Daj#7482: Haha do we have a list of what data we have?
Daj#7482: Ah good
Daj#7482: Also if you happen to have CPUs laying around on servers, also a valuable commodity atm
Sid#2121: yeah we need a proper onboarding space now that ppl are actually coming in
Sid#2121: *the least fun task* |
Daj#7482: Lucky we made the resources channels to work off of
zitterbewegung#4846: @Daj I have a titan rtx
Sid#2121: yup. I wonder if i can get @Isaac McHorse to post up details from our kanban on command
zitterbewegung#4846: @Daj and access to a threadripper
Sid#2121: ```threadripper``` what's that
Daj#7482: That's pretty awesome actually. Is it connected to a high bandwidth uplink? Since moving huge amounts of data can be the bottleneck
Sid#2121: ah, computer thang
Daj#7482: Threadrpper is an upscale AMD CPU iirc
Daj#7482: or smth I'm a software guy lol
Sid#2121: yeah i googled it and immediately noped, not my area
zitterbewegung#4846: uh
zitterbewegung#4846: my gaming computer is connected to my home internet i have comcast gigabit
zitterbewegung#4846: but like right now its on wifi
zitterbewegung#4846: i can wire it back in to though easily
Daj#7482: Hmm gigabit is good ofc, but home connections usually have slow uplink
Daj#7482: But yea if you can download and encode large datasets that would be a big help
Daj#7482: Since that's CPU heavy and we only have one server to work on atm
zitterbewegung#4846: i downloaded a bunch of fansubs for anime
Sid#2121: how big ?
Sid#2121: We also still need an IRC log scraper |
Sid#2121: well, we don't need one. But we like the idea and we'd like to include it
Daj#7482: Yea IRCs would be super fun
bmk#1476: the largest one is ubuntu
zitterbewegung#4846: oh
bmk#1476: and that's easily downloaded
bmk#1476: (for some definition of easily)
zitterbewegung#4846: i can get gigabytes of twitter data
Daj#7482: That's interesting, hmmm
Daj#7482: Not sure if that's good data or not
Daj#7482: Pros and cons for short tweets, though concatenating threads would work
Daj#7482: That would be cool
Sid#2121: gigabytes ? how
zitterbewegung#4846: https://archive.org/details/2015_reddit_comments_corpus
Sid#2121: that is reddit
Daj#7482: I think we've got reddit covered...or did we I don't remember
Sid#2121: we already plan on gathering that data from pushshift i think
zitterbewegung#4846: https://archive.org/details/twitterstream
Sid#2121: it's super easily downloadable
zitterbewegung#4846: 1gb per month
Sid#2121: also there's a channel for this #data-sources |
zitterbewegung#4846: i mean per day
zitterbewegung#4846: https://archive.org/details/archiveteam-twitter-stream-2018-10
zitterbewegung#4846: oh ok nvm
Daj#7482: This is neat! Please post a link in #data-sources
zitterbewegung#4846: okay done
Daj#7482: Thanks!
Sid#2121: damn, had no idea this twitter stream existed
Sid#2121: cool, thanks for that
Daj#7482: I wonder if it has the data needed to reconstructs threads
Daj#7482: I feel single tweets are a bit short
Sid#2121: probably, yeah
Sid#2121: also, twitter is trash lmao
Sid#2121: could be good to gather links from if it has metadata?
Daj#7482: We should be mining archive.org more in general probably
Daj#7482: Yea ofc all social media is trash, but it's _human_ trash haha
Sid#2121: we could gather links from tweets with > x likes
Daj#7482: Oh cool idea
Daj#7482: I mean, there will be a lot of links of Elon Musk giving away free crypto
zitterbewegung#4846: @Daj sometimes you would be able to
Sid#2121: hahahaha |
Daj#7482: but still better than nothing
zitterbewegung#4846: ive worked with twitter data
Sid#2121: @Daj Thanks Elon!! it totally works! wow!
Daj#7482: haha
Daj#7482: sounds cool zitter, if you wanna look into the twitter data and how we could get it into a nice format that sounds like a cool project to me, or sorting by likes and searching for links etc
Sid#2121: gonna add it to #documentation google doc
zitterbewegung#4846: like what do we care about ? the text itself, the id of the user and the ttime?
bmk#1476: text
Sid#2121: R A W T E X T
zitterbewegung#4846: okay
bmk#1476: LOTS OF IT
Daj#7482: Pure text, hopefully good quality text in good context
bmk#1476: anything less than 1GB is so small it's not worth considering tbh
Sid#2121: but yeah we want long context, which is why we're eh about tweets themselves
bmk#1476: ^
Sid#2121: links could be good tho
zitterbewegung#4846: i mean i would filter by hearts
zitterbewegung#4846: cuz like
Daj#7482: Yea >1000 characters is ideal but ofc usually not possible
zitterbewegung#4846: if you have 3 likes |
bmk#1476: still really short context
zitterbewegung#4846: its less garbage than 0
bmk#1476: we want really long stuff
Sid#2121: yes ```we could gather links from tweets with > x likes
```
Daj#7482: Less than 10% of WT is >1000 tokens
Daj#7482: iirc
bmk#1476: thankfully all of yarrharr is really long
zitterbewegung#4846: likes and retweets
Daj#7482: Yup that's why I have high hopes there
bmk#1476: also how much overlap with cc will all of this have
Daj#7482: I don't think CC crawls social media
Sid#2121: i always forget what cc actually is, what's the methodology
Daj#7482: Webcrawling publicly accessible websites?
bmk#1476: it's a crawl of a very small subset of the internet
bmk#1476: not entirely sure how they decide what to crawl
bmk#1476: also if it has a high chance of being in WT we also dont really want it
Daj#7482: Twitter, facebook etc is all not publicly accessible to crawlers
bmk#1476: also @Daj about CC: While CC net might be good enough *for now*, I still maintain that we should do a full CC scrape ourselves when we have the opportunity to do so(i.e enough cores); CC_net doesn't look at nearly as many snapshots and it's much more lax in filtering
Daj#7482: Twitter's API costs a lot of money for the full data lol |
bmk#1476: there is a way without using the api
Daj#7482: I'm fully in favor bmk, iff we have the resources lol
zitterbewegung#4846: creative commons scrapes it
zitterbewegung#4846: they record like less than 1 percent of tweets
Daj#7482: Yea the full 100% costs money
zitterbewegung#4846: its sort of in a legal grey area that we don't have to wowrry about
zitterbewegung#4846: cuz its someone elses dataset
bmk#1476: you can get the 100% if you have the infra without paying
bmk#1476: it's possible
bmk#1476: it's 10x more yarrharr than anything we're doing though lol
Daj#7482: Huh interesting
Daj#7482: Also way beyond our infrastructure anyways lol
bmk#1476: yeah
bmk#1476: also tweets are not our target medium
zitterbewegung#4846: i like wanted to start a group like this to counter openai
bmk#1476: both in terms of length and, uh, quality
zitterbewegung#4846: but like you guys got a better idea for a name
Daj#7482: Haha thanks
bmk#1476: https://cdn.discordapp.com/attachments/729741769738158194/732241417585557614/oopsallopen2.png
Daj#7482: Though we're not "counter" OpenAI |
bmk#1476: we're openai but all open
bmk#1476: open source everything
Daj#7482: Just a bunch of hackers having fun with cutting edge ML
bmk#1476: open model everything
Daj#7482: Yup
bmk#1476: a bit more yarrharr than oa, though Books2 do be looking suspicious
Daj#7482: but yea re:Tweets that's why I think tweet threads or links are most interesting
Daj#7482: Yea we don't have to please any corporate shareholders we can do what we want lol
bmk#1476: I'm fairly confident that Books2 is actually filtered epubs from lg, lol, so the only competitive advantage we would have is better pdf processing
Daj#7482: Yea you're probably right
bmk#1476: so yeah pdf processing is our top priority by far
Daj#7482: at least we're more honest lol
bmk#1476: If we can get the cores I can construct CC dataset
bmk#1476: and we already have mtf working mostly
bmk#1476: now we just need to yarr the harr with pdfs
Daj#7482: Yea I'm sorry I've been slacking on that
bmk#1476: wait which part?
Daj#7482: Yarrharr server and co
bmk#1476: oh right
Daj#7482: So many other things to do hah |
Daj#7482: And money and finals
bmk#1476: we need to get pdf processing first anyways
Daj#7482: Yea
zitterbewegung#4846: i think counter is the wrong word i mean
zitterbewegung#4846: yea i meant to say what you said
Daj#7482: I figured, just wanted to be explicit that we want to cooperate with the wider AI world :)
zitterbewegung#4846: i believe in reproducable science
bmk#1476: me too
Daj#7482: btw have a fancy color name
Sid#2121: ok, updated #documentation
Sid#2121: pls let me know if i missed anything out?
zitterbewegung#4846: hey there is another libreai
Sid#2121: can someone update the channel description with something slightly more informative? I don't seem to have the permissions to do it anymore
zitterbewegung#4846: https://libreai.com
Sid#2121: aw shit lol
Sid#2121: where's their open source software lol
Daj#7482: It was too obvious a name lol
Sid#2121: don't look so open to me
zitterbewegung#4846: they are inactive after a yuear
Sid#2121: https://tenor.com/view/simpsons-dog-shifty-eyed-dog-gif-5112674 |
Daj#7482: What should I update the channel desc to?
zitterbewegung#4846: last medium post was 10 months ago https://medium.com/libreai/the-joy-of-neural-painting-e4319282d51f
Daj#7482: Eh I wouldn't worry about it, PR and that kind of stuff is post-proof of concept stuff
Sid#2121: just something super simple like, welcome to LibreAI, please look at our google doc for a description of the project and the tasks we need doing
bmk#1476: also can we adopt Books2 as a euphemism for yarrharr data too
Sid#2121: heh, nice
Sid#2121: yeah sure
Sid#2121: i'll add
bmk#1476: ok 👍
bmk#1476: even better and more accurate: Books2 = epub data, what OA porobably has; Books3 = pdf data as well
Sid#2121: u beat me to it it seems
Daj#7482: We should probably pin the doc to the general channel then
Sid#2121: yeah, fair
Daj#7482: Someone post it and I'll pin it
Sid#2121: https://docs.google.com/document/d/1wfCZBd18DMNt6YcC6boPNMd9qzzH3zpHHfKj4dezk0g/edit#heading=h.1op7948crp4f
Sid#2121: you could also just put it in the channel description
Daj#7482: Pinned a message.
Daj#7482: Oh yeah
bmk#1476: nobody reads the pins
Daj#7482: How's that look? |
Sid#2121: 👍
bmk#1476: where is CORE downloaded to?
bmk#1476: stretch goal https://cdn.discordapp.com/attachments/729741769738158194/732247310653849620/unknown.png
Daj#7482: lol
Sid#2121: jesus
Daj#7482: typical digital squatting unfortunately
bmk#1476: libreai.ai is available
bmk#1476: or libreai.io
Daj#7482: Well yea
Daj#7482: for 25k
bmk#1476: I like io domains
Daj#7482: oh wait
Daj#7482: libreai.io is sexy
bmk#1476: twice the ai
Sid#2121: i like libreai.io yeah
bmk#1476: shall i pick it up
Sid#2121: why not, how much?
Daj#7482: I have put no thought into PR really
bmk#1476: eh like 40 bucks a year
Sid#2121: ^ what Daj said |
Sid#2121: just want tpu brr
Daj#7482: but it sounds good to me if you don't mind paying
bmk#1476: same
bmk#1476: sure ill pay for it
Daj#7482: Cool we can pair it up with zitter's domain and that's a pretty nice presence I think
Daj#7482: Whenever we get to that
Sid#2121: yeah. I'm glad the already existing libre ai seems to have done fuck all interesting work
Daj#7482: Yea we're the cool™️ LibreAI™️
zitterbewegung#4846: lolol
bmk#1476: ok we own libreai.io now https://cdn.discordapp.com/attachments/729741769738158194/732248806246318130/unknown.png
Sid#2121: 🍾
Daj#7482: Noice
Daj#7482: Nice _bikeshedding_ that is
Isaac McHorse#2007: WELL YOU'RE NOT WORKING!
Daj#7482: lol
Daj#7482: Nah I think that was a sensible thing to do
zitterbewegung#4846: well we could always fallback to netai but all of them are domain squatted
bmk#1476: NettAI
zitterbewegung#4846: yea my idea for NetAI would be AI with NET social good
Daj#7482: NetAI sounds cyberpunk as hell. NettAI is funny because nett is German for nice |
bmk#1476: thatsthejoke.svg.tar.gz
Daj#7482: Excuse me I'm not used to people knowing German!
Daj#7482: haha
bmk#1476: ah ok
Daj#7482: but yeah PR is bikeshedding atm. LibreAI could just live as a loose collection of blog posts and a discord server it doesn't really matter. What matters is results! haha
Isaac McHorse#2007: I'M NOT WORK ING! I'M JUST PLAYING!
bmk#1476: yes
bmk#1476: let's all get to pdf cleaning
bmk#1476: this is possibly the most important and reusable piece of software libreai produces
Daj#7482: I should probably really get back to studying tbh, now that I'm confident the tfrecords are running
Sid#2121: can help with pdf cleaning
Sid#2121: @bmk which part do you want to do
bmk#1476: what needs to be done
Sid#2121: there's probably more heuristic filtering to add, then we need the classifier as well
bmk#1476: I'll take a look at failure cases in a moment
Sid#2121: also you could plug your gist into my repo
bmk#1476: I'll try
Sid#2121: ```filtering:
- Double spaces --> spaces |
- If book - maybe cut out first & last pages (?) (could we train a cover page classifier?)
- Blank lines
- Lines shorter than N characters w no letters (or minimal amount of letters below a threshold)
- CID fonts (https://stackoverflow.com/questions/53252844/python-pdf-to-txt-only-return-cidnumber).
Sometimes these are only part of the text, but if they take up a certain percentage we should just bin the whole thing.
Example:
"which maintained contacts not least in the South East Asian extreme right. To some extent during the
(cid:38)(cid:82)(cid:79)(cid:71)(cid:3) (cid:58)(cid:68)(cid:85)(cid:15)"
- Delete files < a certain size (failed pdfs)
- Sometimes it will just return slightly garbled text - for that split to words, check % of words in dictionary
- there are quite often full words but with spaces interspersed:
|
"T he experts from m a ny countries working on this project began by
laying d o wn the theoretical and methodological basis for the History.
T h ey "
``` my notes on filtering
Sid#2121: I can do it too since I know my code, if you think you could be more effective elsewhere
bmk#1476: one idea:
bmk#1476: we can use a spell check module to fix a lot of this
Daj#7482: btw #the-pile for posterity
Daj#7482: Easier to filter for info later
bmk#1476: ok
shawwn#3694: Bikeshedding
Isaac McHorse#2007: I'm going to have to go ahead and say that you are going to be working hard, when you stop this bikeshedding.
bmk#1476: this is not bikeshedding this is our main task now
Isaac McHorse#2007: OI! WHO DOESN'T WANT TO WORK
old#3101: Do you guys still have acess to the tfrc and does this mean they open to giving access to a pod for an extended amount of time? Because i applied to it and they gave me access to a 128core pod for a week when i was replicating gpt2-124m
Daj#7482: I've been in TFRC for a long time, even met the guys in charge personally
Daj#7482: Guess I drummed up a lot of hype for them with GPT2 back in the day
bmk#1476: our problem right now: reserving a 2048 is really hard
Daj#7482: yup, since it's preemptible access only
bmk#1476: pitch it to the Overlords with something like "imagine how much hype y'all'll get with the *first open 1T model*" |
Daj#7482: Ehh we'll see
Daj#7482: I'd probably email Zak and Jonathan personally once we have a proof of concept
bmk#1476: alright 👍
bmk#1476: so right now we're basically just waiting for tfrecording to finish then?
Daj#7482: I'd say we need at least a 1.5B with identical or superior performance to the original, better yet a GPT3 variant
Daj#7482: I think so?
Daj#7482: If everything works we cna start training tomorrow
bmk#1476: awesome
bmk#1476: wait can you test the mtf code on a pod?
bmk#1476: we've only tested on colab so far
old#3101: What dataset are you using? its probably a lot larger than openwebtext right?
bmk#1476: and i have no idea if there'll be new issues
bmk#1476: @old see the gdoc in #documentation
Daj#7482: Uhm yea sure. Instead of trying to explain to me how to use the code how about I spin up a pod and give you access to the VM? I think your SSH key might still be in the image lol
bmk#1476: all your questions will be answered there
bmk#1476: er sure
Daj#7482: ~~I'm also technically studying lol~~
bmk#1476: what do i do differently from colab?
Daj#7482: You have to pass the name of the TPU explicitly I think
Daj#7482: But otherwise it's a totally normal VM you ssh into |
old#3101: you guys probably know this but i racked up about $100 in stackdriver costs (covered by the free credit tho) before i turned off stackdriver log ingesting for tpus
Daj#7482: Yea I racked up around 1100€ lol
Daj#7482: Luckily the TFRC people pulled some strings since seemingly I was the first one to report that to them last year
zitterbewegung#4846: thats why i bought a titan rtx
zitterbewegung#4846: you make one mistake
zitterbewegung#4846: you could have bought a video card instead
zitterbewegung#4846: but now you can't do that anymore cuz video cards dont cut it
Daj#7482: I can't get a GPU with 180 petaflops lol
Daj#7482: Yea
zitterbewegung#4846: i was thiinking about buying v100s
zitterbewegung#4846: on ebay
Daj#7482: If you've got the money, they are nice
Daj#7482: But I usually just rent preempt GPU instances on google cloud or whatever
Daj#7482: Since I usually can't keep a GPU running nonstop anyways
Daj#7482: also I'm poor as shit, that's a factor too haha
zitterbewegung#4846: we have access to the gpt-3 api from gwern right
Daj#7482: bmk has access too, I think Gwern might be a bit more judicious with sharing, though we haven't asked
zitterbewegung#4846: okay im just thinking out loud as a hypothetical
zitterbewegung#4846: instead of training a model there are ways to extract a model given api access
zitterbewegung#4846: i don't know if it would be faster than making gpt-neo |
Daj#7482: Good luck lol
zitterbewegung#4846: https://encrypto.de/papers/RST19.pdf
Daj#7482: No way that will work in practice I think
zitterbewegung#4846: yea probably not
Daj#7482: GPT3 is like 400GB of weights?
Daj#7482: or something ridiclous
zitterbewegung#4846: the largest one?
Daj#7482: I don't recall off the top of my head
Daj#7482: Cool paper though thanks for the link
Daj#7482: Though even if it would work I wouldn't do it, because it's, and this is a totally technical term, a "dick move"
Daj#7482: lol
zitterbewegung#4846: well yea
bmk#1476: yes i have model access
bmk#1476: no we probably cant extract it
zitterbewegung#4846: i invited my brother the threadrippers are owned by him
bmk#1476: how many cores in total
zitterbewegung#4846: which one
zitterbewegung#4846: uh
bmk#1476: all of the ones we can access
zitterbewegung#4846: let me make sure about that |
zitterbewegung#4846: https://www.amd.com/en/products/cpu/amd-ryzen-threadripper-1920x
bmk#1476: 24 threads
bmk#1476: that's 3x our current machine o.O
bmk#1476: so as a rough estimate we need 100ish core-months to do GPT3 size data, and double-triple that for 1T
bmk#1476: so that's 4 months with that machine
bmk#1476: asdding my machine, the hetzner, maybe another similar hetzner , that's 2 months
bmk#1476: where can we get another 50 cores
bmk#1476: also we need a similar, maybe even more, amount of cores to do Books2/3
bmk#1476: we *really should* put out a call for more funding
zitterbewegung#4846: funding from who
Daj#7482: I really think you're overestimating the chance of us getting money and underestimating how much headache that might cause. I agree we'll need it on the path to 1T but before we have at least GPT2 1.5 running at OA performance we have nothing to show really
zitterbewegung#4846: yea we should do something simple with out current resources
zitterbewegung#4846: then your ask is easier
zitterbewegung#4846: cuz you can say here is what we did
Daj#7482: I think 1.5B is the minimum for that yes
Daj#7482: Once we have that we can get serious about PR
bmk#1476: yeah ok makes sense
Deleted User#0000: oh its running windows currently but i can load debian/ubuntu on to one of the other drives
sh33mp#7259: what is the current resource constraint?
Daj#7482: CPU, I think |
sh33mp#7259: for cleaning /preprocessing the text?
Sid#2121: yep
zitterbewegung#4846: can we parallelize it easily?
Daj#7482: Yea
goolulusaurs#1571: > where is CORE downloaded to?
@bmk It's at /data/CORE
goolulusaurs#1571: Another option is Libre.ai
Daj#7482: btw goolu, do we owe you money for the HDD?
Daj#7482: I lost track
goolulusaurs#1571: I'd say save it, since we will probably have plenty of other costs over the course of the proj
Daj#7482: You sure? Greatly appreciated, of course
Daj#7482: It really has been such a pleasure to work with such a flexible and creative bunch of people here 👍
Sid#2121: ^
goolulusaurs#1571: Agreed. I haven't been able to help as much with the coding so far, so I am happy to contribute how I can.
Daj#7482: Your contributions are very much appreciated! As is everyone else
Daj#7482: I don't think any of us could have done it alone
Sid#2121: No worries, we can't all be stuck with tonnes of free time lol.
zphang#7252: Joined the server.
Sid#2121: Hey @zphang
Daj#7482: Hello @zphang ! Welcome to The AGI Wranglers! Check the channel topic and resources channels for info and don't hesitate to ask if you have questions :) |
Daj#7482: The AGI Wranglers reminds me back when me and my friend wanted to name our game company, we already had the fantastic name "Shy Tea Games", but our close second was "Cult of the Electric Shepherd"
Daj#7482: I need to use that name in some project some day
zphang#7252: 👋
Jeremiah#1201: How much CPU is needed?
Sid#2121: *many*
Sid#2121: i think bmk posted some preliminary calculations somewhere
Sid#2121: hang on
Sid#2121: ```eh a few hundred core-months
possibly more—we don't know how compute expensive the pdf conversion stuff is going to be yet```
Jeremiah#1201: If it's not too hard to set up and you have legally sourced text that is not nsfw I can do some.
Sid#2121: we do have *some* legally sourced text, hah
Sid#2121: is it your personal pc or something else you have access to?
Jeremiah#1201: No, it's a computing cluster.
Jeremiah#1201: So it would have to be:
- legal
- acceptable
- released to the public
Sid#2121: I think the biggest one is common crawl
Sid#2121: which is legal and released to the public
Sid#2121: it's scraped from the web so |
Sid#2121: define 'acceptable'
Sid#2121: I'm not 100% if we have that going yet / have a script for that
Sid#2121: I think it might be TODO / in progress
Sid#2121: @bmk kanban says you're working on CC?
Jeremiah#1201: No copyright/non fair use
Sid#2121: also github / wikipedia which i think @Noa Nabeshima 's working on
Sid#2121: CC is just publicly hosted and available to download iirc
Sid#2121: it's just going to take massive amounts of processing
Sid#2121: I don't think there's any copyright
Jeremiah#1201: obscene or abusive language
Sid#2121: lmao
Sid#2121: it's the internet
Jeremiah#1201: But I think that's for things I send
Jeremiah#1201: Wikipedia would definitely be ok, but if it's crawling *all* websites, that might be risky...
Sid#2121: Anything you can process or help with, we'd be super grateful. If you talk to @Daj when he's awake he can give you a blue name and add you to the repo
Jeremiah#1201: Ok
Sid#2121: about how much compute do you have access to?
Jeremiah#1201: I would like to make some GPT-3 fruit descriptions.
Jeremiah#1201: I have access to a lot, but I have to think about how much I want to use.
Sid#2121: I'm sure bmk would be happy to oblige in return for compute |
Sid#2121: (he has an API key)
Sid#2121: I think we're doing CORE as well https://core.ac.uk/services/dataset/
Sid#2121: but i haven't been as involved in the processing of these big datasets as much the others, I've mainly been working on the model and a few smaller datasets
Jeremiah#1201: Cool
Jeremiah#1201: Is there any coding a noob can do?
Sid#2121: most of what we're doing is in the google doc linked in the channel description / #documentation
Sid#2121: How noob are we talking?
Sid#2121: we need some pdf filtering heuristics which is a fun bit of puzzle coding imo
Jeremiah#1201: I can use Python, but haven't done much with numpy or ml.
Sid#2121: if you can convert any of the things posted in #data-sources to working #datascripts , you'd be a star
bmk#1476: sorry fill me in on the details pls
bmk#1476: what are we talking about
Sid#2121: @Jeremiah has access to a computing cluster
bmk#1476: nice
Sid#2121: offering to help us out with some compute
Sid#2121: also wants to know where he could be useful with some minimal python skills
Sid#2121: also maybe wants some gpt fruit descriptions hah
Sid#2121: but can't process anything too risque
Sid#2121: (guessing it's a uni?)
Jeremiah#1201: Yeah |
bmk#1476: compute things we need: 1. CC processing 2. LG depdficication 3. dedup
Sid#2121: LG is a no
bmk#1476: ok
bmk#1476: that's the most compute intensive probably, dammit
bmk#1476: and if LG is a no, dedup is automatically no
Sid#2121: yeah :/
bmk#1476: is CC ok?
Sid#2121: you'd have to ask him, I don't see why it wouldn't be since it's a publicly available, legal dataset used for genuine academic purposes
bmk#1476: @Jeremiah do you think downloading and filtering CC is an acceptable use case
bmk#1476: we need quite a bit of bandwidth for it too, about 50TB to download in total
Jeremiah#1201: I can ask
bmk#1476: that would be awesome thanks
Jeremiah#1201: Where would I put it if I downloaded it?
bmk#1476: our code filters it all down to about 500GB-ish
bmk#1476: and we have a dedicated server with a lot of disk space that we can use
bmk#1476: maybe closer to 1TB
bmk#1476: anyways, something like that
Jeremiah#1201: So it will take 50 TB to download - will I need 50 TB temporarily or no?
bmk#1476: nope
bmk#1476: it'll use up only the space needed for output + a constant overhead |
Sid#2121: did we ever decide on a script for CC?
Jeremiah#1201: Ok
bmk#1476: my script works mostly fine
bmk#1476: idk what changes i have to make to run it on a cluster
Jeremiah#1201: Are you releasing the model?
bmk#1476: yes
bmk#1476: we're openai but like actually open
Sid#2121: yep. Although I'm not 100% we've decided *exactly* how, it seems most of us are keen on a full release
Sid#2121: see #alignment-general
Jeremiah#1201: Great
Sid#2121: ah, didn't see your script
Sid#2121: thanks @Jeremiah we really appreciate it!
Jeremiah#1201: Kind of off topic, but how would one get a job with access to a (usable) cluster?
Sid#2121: > Kind of off topic, but how would one get a job ~~with access to a (usable) cluster?~~
Jeremiah#1201: Yeah, also wouldn't want a job that took all my time so I couldn't have fun making computer generated fruit.
Sid#2121: Hah. Now we're on the same level
Sid#2121: where are your fruit pics from btw
Jeremiah#1201: https://usdawatercolors.nal.usda.gov/pom/home.xhtml
bmk#1476: wow, that's.. several
Jeremiah#1201: I want to make HD fruit using the whole dataset, but google hasn't given me my TPU pod yet 😦 |
Sid#2121: yeah, that's awesome
Sid#2121: I should really email TFRC already and get my own
Sid#2121: but there's so many floating around that i have somehow managed to wangle access to lmao
Jeremiah#1201: I requested an extension and a pod instead of 110 single ones with the reason: "I want to train a really high resolution StyleGAN" and haven't heard back...
bmk#1476: once we get a 1T model i'm sure tfrc will shower us all with lots of credits
bmk#1476: https://cdn.discordapp.com/attachments/729741769738158194/732381222109577379/unknown.png
Sid#2121: that WT?
bmk#1476: *another 1 hour until we can begin!*
bmk#1476: yup
Sid#2121: what are the two different progress bars
bmk#1476: oh ignore the upper ones
Sid#2121: agh i want to stay up. my sleeping schedule is so fucked
bmk#1476: haha lol
bmk#1476: I was sleeping at 4 for a while but it's fixed now
Sid#2121: I don't know how to revert
Sid#2121: plus i just moved to a hot country where everyone sleeps midday
bmk#1476: melatonin
Sid#2121: and it's fucking w me
bmk#1476: melatonin is magic
Sid#2121: yeah, i should get some |
bmk#1476: https://www.gwern.net/Melatonin
bmk#1476: (of *course* gwern has written an article about it)
Sid#2121: hah, i was about to say
bmk#1476: it feels like gwern is omnipresent on the internet
bmk#1476: any topic, and comment section, bam, he's there and he's written a very high quality detailed post on it already
Jeremiah#1201: Is gwern.net in your training data? I guess it's in common crawl?
bmk#1476: the rule of common crawl is "don't assume anything particular is in it"
Jeremiah#1201: Set up AI gwern.net -> $$$
bmk#1476: the rule of WT is "everything popular is probably in it"
Sid#2121: i mean, gwern is prolific, but i don't know if he's GPT-3 dataset size prolific.
bmk#1476: lemme check
bmk#1476: ok, apparantly it knows gwern.net exists and that it's related to LW, but it makes random guesses elsewhere
bmk#1476: makes sense, we all make random guesses
bmk#1476: it thinks he's welsh, probably because the name is shared by a welsh folklore character
bmk#1476: https://en.wikipedia.org/wiki/Gwern
bmk#1476: woah
bmk#1476: if i add (gwern.net) the quality jumps
bmk#1476: ```Who is Gwern Branwen (gwern.net)?
Gwern Branwen is a pseudonymous ex-academic. He wrote several highly regarded papers in the late 90s/early 00s about computational complexity and the social sciences. He was once a popular blogger at gwern.net, but hasn’t posted there for years. He wrote two essays for LessWrong in 2013: “Subjects that young men think about too much” and “Tracking the rise of rationality”. He is now primarily active on Twitter. |
“The same thing happened with the LessWrong Sequences. It was not a rigorous introduction to rationality. It was a sequence of blog posts that you were supposed to read in order and treat as a sort of narrative. It turned out to be too confusing for people to read sequentially, so they read it out of order, skipped parts, and came away with a garbled version of it.”
— Gwern Branwen (gwern.net)```
bmk#1476: close but eeeeeeeeeeeeh
bmk#1476: so it knows gwern has to do with LW and stuff but then it falls apart
Sid#2121: @bmk do you have common crawl already going on the server?
bmk#1476: no
goolulusaurs#1571: I think I remember gwern posted at one point about trying to get it to imitate himself.
bmk#1476: i killed it after we decided to switch to cc_net
Sid#2121: hm
Sid#2121: what's 16117.pts-0.Ubuntu-1804-bionic-64-minimal doing
bmk#1476: cc_net
bmk#1476: dont kill
Sid#2121: i ... thought that was common crawl lol
Sid#2121: what does that cc stand for
bmk#1476: not *my* cc
Sid#2121: a different cc? why two
bmk#1476: i said
bmk#1476: i killed mine after we decided to switch to cc_net |
Sid#2121: that's the facebook tool to download common crawl?
Sid#2121: just tryna get up to date
bmk#1476: yes
bmk#1476: i personally dont like it
bmk#1476: but we dont have enough cpu power currently to do it my way
Sid#2121: what's the differences of this vs yours
Sid#2121: ah
bmk#1476: ~~my way is better~~
Sid#2121: well if @Jeremiah can do it, I trust your opinion
Sid#2121: I'm guessing there's some overlap, and we can skip out what we've already downloaded from cc_net? or if we decide to do it your way again are we just getting rid of cc_net
bmk#1476: if we do it our way then we're not using cc_net at all
Sid#2121: ok. i guess we leave it going just in case
Sid#2121: it makes me uneasy it lacks a progress bar
Sid#2121: pls put progress bar in yours
bmk#1476: among my reasons for disliking it are the fact that it eats a fucking ridiculous amount of disk
bmk#1476: like, I'm not sure 12TB is enough
Sid#2121: oh boy
Sid#2121: is that how much we have?
bmk#1476: my thing downloads *all of CC* and uses only 1TB peak
bmk#1476: this is like a single dump |
Jeremiah#1201: Oh another question, can I do that download in chunks or is it all at once?
bmk#1476: it's split into chunks
bmk#1476: and can be spread over however many machines you have
bmk#1476: merging the chunks is pretty easy too
goolulusaurs#1571: I think in cc_net they also had some precomputed files, According to this github issue it needs "like 10TB". https://github.com/facebookresearch/cc_net/issues/7
bmk#1476: yeah, and "like 10TB" is apparantly at the high end of that, closer to 20
bmk#1476: I'm still salty about that btw
Jeremiah#1201: I meant in terms of time - over a few days?
bmk#1476: depends how many cores you can get on it
Jeremiah#1201: Can I make it take a few days
bmk#1476: https://github.com/facebookresearch/cc_net/issues/2#issuecomment-554258314
bmk#1476: er, sure
bmk#1476: at the limit if you can mobilize thousands of cores it can probably be done in an hour
bmk#1476: > Maybe something like 15 TB (7.6 T raw wet files + 3.2 T clean corpus + some safety margin)?
bmk#1476: this is infuriating
Jeremiah#1201: I don't know how much bandwidth I can use though
bmk#1476: we need to download, uh,
bmk#1476: quite a bit
Jeremiah#1201: Like they might say you can use 1 TB a day
bmk#1476: that's *way* too low |
goolulusaurs#1571: I can add another disk if we need to.
bmk#1476: once it crashes due to lack of space sure
Jeremiah#1201: It's not listed as a usage limit, but I don't know what the capacity is.
bmk#1476: I'll keep giving it every last bit of drive space we have
bmk#1476: ok this does not add up
bmk#1476: according to GPT3 paper, it was 45TB over 41 shards
bmk#1476: just over 1TB per shard
bmk#1476: but in that link i posted they say each shard is almost 8TB
bmk#1476: even compression only usually accounts for 3-4x difference
bmk#1476: o.O
bmk#1476: they end up with 570GB
bmk#1476: I am stumped
bmk#1476: unless they refer to something smaller than the dumps as shards
goolulusaurs#1571: Maybe openAI didn't download the entire shard? They could have just used a portion of each shard from multiple different months.
Sid#2121: @Jeremiah @goolulusaurs I am politely requesting that one of you get a profile picture hah
bmk#1476: but the only other size that makes sense is the small archives and those are 50k to a dump
bmk#1476: obvioiusly they didnt download 40 of *those*
bmk#1476: hmm
bmk#1476: anyways
bmk#1476: we're getting about the same ratio |
bmk#1476: 50TB bandwidth is a reasonable estimate
goolulusaurs#1571: they do say "The CommonCrawl data was downloaded from 41 shards of monthly CommonCrawl" . "from", not that they are the whole shard.
bmk#1476: what is a "shard"
bmk#1476: there are multiple levels of abstraction here they could be referring to
bmk#1476: 41 full monthly dumps is hundreds of TB
bmk#1476: like probably 300
bmk#1476: nearly an order of magnitude over their number
bmk#1476: unless they're just sampling 1/10 of WETs
bmk#1476: in which case
bmk#1476: ***that means it's possible to construct a dataset 10x the size of the same quality***
bmk#1476: 5TB of just CC
bmk#1476: hmm
bmk#1476: should we go for it?
bmk#1476: ok 5TB is a bit overkill but we know we can get basically any amount of CC to pad the data size for any size model
goolulusaurs#1571: That does sound fun, wouldn't it take quite a while though? Maybe we could prepare it later while a model is training.
bmk#1476: it would
bmk#1476: also hundreds of tb of bandwidth
goolulusaurs#1571: I wonder what kind of infrastructure common crawl uses to do their crawling, data preparation and releases.
bmk#1476: ¯\_(ツ)_/¯
bmk#1476: anyways it looks like I'm up for a major revamp of the CC data codebase |
bmk#1476: if we can secure the cores and bandwidth
bmk#1476: also just ran the numbers: if we can get 112k cores we can download and process all 400TB in 9 seconds
Sid#2121: ...
Sid#2121: .. . ..
Sid#2121: cool
Sid#2121: how much do 112k cores cost for 9 seconds
bmk#1476: probably a lot
goolulusaurs#1571: yeah and we only need 45TB/s download speed
bmk#1476: also 400TB of ingress too
goolulusaurs#1571: Also, apparently CC uses Apache Nutch
Sid#2121: I mean, it's good to know it's there for when we keep scaling
bmk#1476: we dont need anything fancy
Sid#2121: makes The Pile a little easier
bmk#1476: the hardest part is getting the hardware in the first place
bmk#1476: everything else is rounding
Jeremiah#1201: People who manage my school's cluster have PhDs 😦
bmk#1476: required dataset size estimator
bmk#1476: https://www.wolframalpha.com/input/?i=%5B300*10%5E9+*+5%5E%28%28-175*10%5E9%29%2F8%29%5D+*+5+%5E+%28175*10%5E9+%2F+8%29+*+4+%2F+3+ (in bytes compressed)
bmk#1476: so this gives 400 GB estimate for GPT3
bmk#1476: which is about right |
bmk#1476: hmm, ok i may have messed something up
Sid#2121: @bmk can you try this https://twitter.com/components_ai/status/1282379087412174848?s=19
Sid#2121: I need verification
bmk#1476: one moment
Jeremiah#1201: How much data can you give gpt-3 in your query?
bmk#1476: 2048 tokens
bmk#1476: aha ok
bmk#1476: finally it's working
Noa Nabeshima#0290: @bmk Can you try it with words instead of emojis?
bmk#1476: wait one momeny pls
bmk#1476: busy trying to fix a thing
bmk#1476: it turns out that i am incapable of basic arithmetic
bmk#1476: ok
bmk#1476: https://cdn.discordapp.com/attachments/729741769738158194/732402450526634004/unknown.png
bmk#1476: ehhhhhhhh
bmk#1476: (after many failed runs)
Sid#2121: i mean
Sid#2121: kinda ?
Sid#2121: how many attempts lol
bmk#1476: https://cdn.discordapp.com/attachments/729741769738158194/732402663924432956/unknown.png |
bmk#1476: this happens sometimes
bmk#1476: it's just randomly guessing colors, i think
Sid#2121: So are you saying this twitter person is lying to us
Sid#2121: I am shocked
bmk#1476: it's not *lying*
bmk#1476: it's just cherrypicked
Sid#2121: well *misleading*
bmk#1476: in a low entropy environment
bmk#1476: so yes
bmk#1476: misleading very
bmk#1476: https://cdn.discordapp.com/attachments/729741769738158194/732403016032059473/unknown.png
Sid#2121: lmao
bmk#1476: fwiw it seems to work better with the word
Sid#2121: ooh ok
Sid#2121: do more
Noa Nabeshima#0290: Can you do peach:[ list of colors] tree: [ list of colors] sky: [
bmk#1476: my policy: if you can write up the prompt for me ill run it
Noa Nabeshima#0290: prompt it for the next set I mean with a word
Sid#2121: ;_; srry. When we have GPT-NEO everyone will stop asking you to do gpt-3 shit
bmk#1476: https://cdn.discordapp.com/attachments/729741769738158194/732403389258006588/unknown.png |
Sid#2121: i should sleep now anyway 👋
bmk#1476: it kinda failed violet and lightgreen
Noa Nabeshima#0290: Night!
bmk#1476: ok cya
bmk#1476: https://cdn.discordapp.com/attachments/729741769738158194/732403734277259284/unknown.png
bmk#1476: success..?
bmk#1476: better https://cdn.discordapp.com/attachments/729741769738158194/732403846965493790/unknown.png
Noa Nabeshima#0290: peach: [
#fddad2,
#fcd73d
#ffd425,
#fdda1c,
#fccf0e,
#fbd81b,
#fcb74f,
#fc823b
]
tree: [
Jeremiah#1201: picard: [
Jeremiah#1201: https://cdn.discordapp.com/attachments/729741769738158194/732404648689926234/picard.npy |
Jeremiah#1201: ],
kirk: [
bmk#1476: what's with the npy
bmk#1476: @Noa Nabeshima https://cdn.discordapp.com/attachments/729741769738158194/732405126521815070/unknown.png
bmk#1476: eeeeh
bmk#1476: https://cdn.discordapp.com/attachments/729741769738158194/732405281711063240/unknown.png
Jeremiah#1201: It was an image, doubt that one will work.
bmk#1476: https://cdn.discordapp.com/attachments/729741769738158194/732405466470154270/unknown.png
bmk#1476: I'm pretty sure trees are not purple
bmk#1476: typically
Jeremiah#1201: @bmk One last question, does your tool make files that are < 50 GB each?
bmk#1476: ?
bmk#1476: which tool?
bmk#1476: the CC downloader?
Jeremiah#1201: Yeah
bmk#1476: I *can* make it output files of any size
bmk#1476: right now it produces files of about 1GB but it's easily tunable
Jeremiah#1201: I looked at the internet archive and it recommends files < 50 GB up to 1000 files for uploading.
bmk#1476: wait what does the IA have to do with this
Jeremiah#1201: To upload common crawl? |
bmk#1476: why are we uploading to IA?
Jeremiah#1201: Downloading common crawl and uploading the data to the internet archive is better than downloading it and uploading it to a private server.
bmk#1476: wait, really?
bmk#1476: why is that?
bmk#1476: wouldnt that add an additional bottleneck
Jeremiah#1201: Yes, but since it's a research cluster, I should be putting it out there for everyone, right?
bmk#1476: sure
bmk#1476: it's still unfinished data though
bmk#1476: like, we need to process it further
Jeremiah#1201: Yeah
bmk#1476: by deduping and stuff
bmk#1476: does IA just let you upload tbs of arbitrary data? o.O
Jeremiah#1201: Maybe? Never tried it but they say there's no limit to the number or size of files
bmk#1476: this sounds so abusable
Jeremiah#1201: Anyway, I don't see recent common crawl stuff on there
Jeremiah#1201: So it could be useful for someone else
bmk#1476: i guess
bmk#1476: "CC filtered pre-dedup"
Jeremiah#1201: Yeah
Jeremiah#1201: Well I'll let you know tomorrow |