Видео

We missed you! See ya next week 🫡

Hey Michael, thanks for the great paper! We'll be doing a part 2 next Friday, Oct 18th live at 10pm PST if you have any interest in joining then. We've had a few authors join in the past and those have been fan favorites. We also have a bunch of folks in our Discord who I am sure would love to ask questions if you introduced yourself. discord.gg/s3tBEn7Ptg

Thank you! We are using a combination of Notion for the notes and OBS for the video

We did do a practical implementation blog post and video here: www.oxen.ai/blog/how-to-train-diffusion-for-text-from-scratch Let us know what you think!

This is good intuition 💡 thank you!

YES! I have been saying for 2 years XML is the BEST way to serialize in LLMs because of closing tags

Also it can semantically handle deeply nested graphs as linear tokens. Any of the overhead of "prettyfying" can be done post process trivially taking out all the headaches of do I need 1 space or 3 or 4 or 8 or 15. Zero is the right answer I need zero spaces for the new object I just need a closing tag and move on.

Completely agree, despite @OpenAI pushing structured json format, I used something (I didn't knew at the time) similar to xml, bra-ket notation, used in quantum mechanics, much more from tinkering than actually the formal aspect. Bra-ket is is something like <a|H|b>. it became more involved than that as I developed this in prompt programming in GPT, based on mathematical results I have for 2 years, that I got from a research I did (play I mean, nobody paid me anything, it was just that API was free and as a mathematical physicist I'm used to MATLAB). I am also ADHD, had this idea of fixing chatGPT 3.5 ADHD and give her a friend. It wasn't hard, I mean it is brutally hard, but coincidentally GPTs, all chatbots really, falls as a trivial case of the area of my research. I didn't knew all the bus about AI, terminology etc. IT are not famous for rigour, mathematical methods and specially respect for like, giving credits where it's due, citations. I stumbled on some papers and other videos and finally discovered where Transformer, loss, gradient descent comes from. All these things are well know, like linear transforms, steepest descent and relative absolute error. It really makes one head hurts so much low bar coming from these so called godfathers of AI. Offense intended, they create these things, they're not formally Models, let's not get into Tensors. I'm pretty sure Yllia sucks and their peers don't really know tensorial calculations, they did of course General Relativity would be easy. For sure everyone here is well versed in quantum mechanics, thermodynamics, noether theorem, universality classic, philosophy and deep connection with semantics, languages in general, are bilingual. I guess not. Well I actually already solved the alignment problem of LLMs using causality instead of data. Other aspect is is no AGI is necessary, all stochastic character emitters fall in this class.. This means security is a solved problem, but llm code modification by human, me in the case, just as easy. I do every day, don't understand Python, that maybe because NumPy and more I know from inside, the algorithms, also, if history and actually highly intellectual persons agree, is 1. Most people are not brilliant, rarely there's much wisdom in common reason. Object oriented language, python in special is so common because... it's made for glorified typewriters, coders. Have anyone created algorithms? guess not. Has anyone questioned why pytorch and tensorflow are free? guess not. Programmers are pursuing their disposal, surely you won't be missed. Hacking these systems, this job will the only one left.

Thanks for the feedback! The export of the presenter video for zoom was pretty low quality since we do it live. Will see if we can get a higher quality one next time!

Thank you we do our best! Let us know if there are any papers you would want us to cover in the future ❤️

🤜 🤛

Hi... how do you think we can add eval_dataset to the trainer? The **datamodule , passed onto the trainer has eval_dataset as None

@@adhilaseem2518 We have another follow up blog post here where we tried it on some real data: www.oxen.ai/blog/fine-tuning-llama-3-in-14-minutes-using-reft Let me know if it helps!

@@oxen-ai hey, I enjoyed reading your blog. I am facing an issue as my task is not classification, but extracting certain headings from a given document. Since the answer/output does not fall into predefined categories, I cannot use accuracy, precision and recall. What metric do you think I should evaluate for my task? I tried to add eval_dataset to the data module so that I get an idea of the cross-entropy loss on the validation set. But it's giving me some errors. It would be really helpful if you can tell me how this can be done in the right way. I think I am preparing the eval_dataset incorrectly!

You’re welcome! Feel free to suggest any papers you’d like us to cover in our discord as well

Hey there! We added all the notes to our blog here as well: www.oxen.ai/blog/arxiv-dives-fast-speech-2

Oh wow, that is a fun fact

🔥

We try to contact the authors of the papers we cover so hopefully next time!

That’s dope, mind if I ask the use case for seq len 256? I’m curious what the dataset looks like

@@oxen-ai well I figure when you’re pretraining, u just need to make it good enough to spit out semi normal sentences, then the rest can be ironed out thru finetuning. You can introduce rope during finetuning as well to increase your context length / max position embeddings

@@oxen-ai and I also usually pretrain on the dataset I’ll fine tune it with (it seems to help everything stick better) the only difference in pretrain / finetune is the finetuning dataset is in alpaca format, the pretrain dataset doesn’t have the alpaca prompt format just each column in a single text field. Hope this helps

@@spencerfunk6697 Ahh that makes total sense. Smart!

Appreciate it!! Let me know if there are any other topics you want to dive into

Awesome! Excited to see you additions!

Thanks for the question! We have a discord would love to continue the conversation there. A bunch of smart people who can chime in as well discord.gg/s3tBEn7Ptg

RLHF by itself doesn't need labeling but the reward model (that RLHF is based on) needs preference data to be trained. In case of DPO, rather than using preference data to train a reward model, we use preference data to train the final LM itself.

In theory you could train these neural networks to approximate any data, I'm curious what you think the dataset for that would look like!

There were some issues with the edit and we had to re-upload unfortunately :(

@@oxen-ai oh thank goodness it was jjust that. so i was curious, could you finetune a qlora adapter this way?

@@spencerfunk6697 Yep should be able to since it is simply a transformer under the hood! That would be fun to try with the pre-trained ones provided.

@@oxen-ai please make a vid 🙏🙏🙏

Yes, I love this idea. I think someone should try it and see how well it works. We dived a little into the code in our next video as a jumping off point!