Just awesome! I never expected to encounter a fine-tuning video that had such a sweet human touch at the end :-). Keep up the great work and good luck with your venture!

I didn't know it was physically possible to smile this much when debugging . My whole world has been rocked (dead serious). Thanks for the great tutorial!

I tried setting up the Brev instance. I’m already in love with the ease of setup & SPOT pricing 🙌🏻

Wonderful video with a good level of detail and explanation (not just choppy soundbites). Hope you can make more. Many thanks!

your explanation of rank and alpha was the most clear and concise I’ve heard. 🎉 tried reading about it and it’s always been a jumble.

Honestly never seen coders make programming look so fun.

Pretty lucid explanation of the stuff involved. Thumbs up.

I came here for a fine-tuning tutorial and I stayed for Harper ❤ cutest tutorial ever!

Yeah, I agree with other comments, very sweet. Your smile is contagious. I wanted to learn about fine-tuning, but found myself smiling lol. Two birds with one stone! Thank you.

Awesome tutorial! Nice that you are explaining some of the things others would skip over (r & alpha for instance). Will try to teach Mistral understanding a new language (Danish) that way. Let's see how much it will take on 2 A10s.

My god that output was adorable. Don't worry, I have plenty of journal entries just like that haha. Awesome vid.

Hell yeah, can't wait to try brev

please make a notebook for Llama 3 and this time show how the dataset example is formatted. the entire world would appreciated. Love you and your work.

good shit gang, very useful

Great video.

Awesome video, thanks!

thanks! i've been trying to find the correct format for the mistral datasets but i could only find very brief or technical information and the notebooks explaining how to train on your own dataset are almost non-existent. they all assume you intend to download the dataset from hf or use the vanilla models from HF. you've saved my life!

Very cool and pragmatic tutorial. Inspired me to install and train my own M7b model. Would be great to share more video in this topic. Personally I would be curious to know more about different fine tuning methods and how to turn raw information to training data. Thanks for the effort and congrats for the content!

This startup is so cozy

You are indeed correct! FSDP is not needed for training with just a single GPU (assuming your model fits on a single GPU). It is a distributed training method which you use to accelerate your training.

el mejor y mas agradable tutorial que he visto en los ultimos 700 años 😁😁😁😁😁

Honestly, I have never seen a tutorial made with so much positivity in my life before! You have a great personality for this stuff. I wish you and your partner the best with brev. God bless, Jesus love you.

I'm a bit confused on preparing the data. If I was to have a plain .txt file of a journal entry is all I need to do to prepare it just convert it to .jsonl? So for example I could have 150 entries of separate journal entries?

Awesome video!! Thanks for sharing. I'm new to this field. Do you have any experience you could share regarding fine-tuning with structured data such as database tables?

Hey, you have got me curious, been for a looksee through brev and keen to learn more of your offerings. Are you more active on some other platform or just KZfaq, guess I can find for myself but just being lazy. keen to hear more and especially what your future trajectory looks like for open source enthusiasts, mmmm or otherwise. Your a laugh a minute, don't edit that out ffs! Some laughs and a bit of dance is absolutely golden, a brief refresh and back into it is great! Can't be oh so serious all the time otherwise we wind up like the Joker😆

There are two versions of Mistral 7B. I dont know which one to use for my task: "Fine tuning the math dataset in order to improve the capability of answer elementary math questions". "mistralai/Mistral-7B-Instruct-v0.1" "mistralai/Mistral-7B-v0.1 Could you please tell me know. Thank you

Very nice vidéo Thx .

Thank you for this tutorial espiecially the notebook file. Let's say I would like to train the model with my own data related to my work. Let's say I will prepare series of questions and answers . How to prepare eval_dataset file? Can I use the same content as in train_dataset?

5:18 best part :)

The format for the output in JSON. Very popular format for data input/output and what most sites uses for backend calls. Not surprising you'd put it in this format.

Can you show examples of samples for training and eval data?

Can I use it on my m1 Mac pro ? Without any additional GPU thingy?

can i know how do we evaluate after finetuning the model, pls make a video of that

is it possible to convert this new model to a GGUF file so that i can use it with ollama locally?

Just wonder, is there any functions or platforms can help with the creation of my own dataset of local files?

Just a little warning: When you split your set into two sets namely train and validation and skip the test set, in a possible future scenario you would be treating your validation set as your test set (because you did not split into 3 at the beginning). And this is a problem. Although the validation set is not used to change the model weights and thereby caused no data leakage, you're using the losses of validation data to change your hyperparameters such as number of layers, dropout etc. So what you're doing is trying to get the best result based on the losses of your validation set. Do you see what's happening? There's no data leakage, but you're imposing a leakage of a different kind, say, hyperparameter leakage (I just made that up). In a possible future scenario, your test set would be the one you used to tune your hyperparameters. The model could be performing nice with those data and hyperparameters combination but who can give a guarantee it performs not that good with different combination?

11:50 Hey, I was wondering if the max_length can be any random multiple of 2 or is it restricted to a specific number? Lets say I have a dataset where each sample/note is 10k-15k tokens long. Can I set the max_length to 16384 in that case and expect the model to give decent results provided that I train it on enough data? I'd also like the model to have a good context window (preferrably 32k)

Could you provide more information on why you chose to do padding = "left" because most examples seem to be using right padding. Also, it would help if you provided general guidelines on setting up the prompt for cases where we have more than 1 input along with instruction and a corresponding model response. For example, if we wanted the model to generate a response to an instruction like - create a short poem on topic "" in "". So while the instruction is common, the inputs and model response are say, 3 separate columns in the fine-tuning dataset. I hope I am making sense :)

Excellent tutorial. I have numerous text files. How can I convert the content into a Q&A dataset? Are there any free and open-source converters available? Or is it necessary for a human to manually read through each text file and create the Q&A data one by one? Are there any specific guidelines for manually creating Q&A data?

Can you please share notes.jsonl and evaluation.jsonl files?

hello, please do a video how to make rag on zephyr gptq. all gpu poors need this. thank you!

Can you explain the step with the dataset again. You only told you will not show because it is private, but a sample would me nice. So you are not creating questions / answer pairs?

is the process same for llama 2 as well?

I don't even understand why there are notes_validation.jsonl. As no example of notes were shown in this video it is hard to understand what is what for a beginner. I imagined something like you just feed with notes that you want to save and just train on that data to respond accordingly. What is validation data? Is it like if I saved to my iPhone notes that sun is yellow and first entry in notes.jsonl is "The sun is yellow" but notes_validation.jsonl? Do I have to put the same entry in notes_validation.jsonl "The sun is yellow"?

Amazing! Can this work on books? Did you try it?

Where do I get the notes.jsonl file?

Excellant tutorial I have a pdfs data i just want to train a question answer model on my pdf's. questions are common but answers will be vary as per pdf changes Also how to make dataset to train that type of model. how can i develop by using mistral Explain in detail

Ohhhh she got that jungle fever.

i dont have any background in CS and need some help and guidance if anyone is willing to support me. can this work in other formats, for example: { "source": "social media", "input": [ "Thank you " ], "output": [ "I told him, he is in the chat Rn " ], "response_time": 60.0 (

Excited to watch this while I eat lunch later 🤌

oh yeah, tell them to get you a better mic or/and and driver. So tomorrow at the Brev Morning meeting (if you have one) be sure to remember that "hey I need a better mic, I got given shit on KZfaq for having shitty sound, other than that they loved it". Looking forward to the next installment!

Hello, there is a channel called "1littlecoder" that has reuploaded your video, I'm wondering if he had permission to do so

Hey, I am very passionate about llms and wanted to assemble a dataset of my own writings and that of my enlightened gurus. I am technically unsophesticated for this task so I request you to help me out. Will you please train a model in a similar fashion on the dataset that I provide you? (like how you do in this tutorial)

this was so silly. 604 more comments and we will have reached our goal!

Hello, I find your tutorial very interesting but I don't understand everything... my English, sorry, leaves something to be desired! So if you're looking for an idea to improve your tutorial, it would be nice to activate the subtitles, because when I read it's already much better. If you also activated the other languages that would be great! You understand, this isn't a criticism in any way, it's just that I want to make sure I've understood correctly. Otherwise, you probably know that the Mistral is a strong wind from the South of France... The old Provençals say it's a wind that drives you crazy and lasts 3, 6 or 9 days. Have a nice weekend

Wow they only have 4K subs? Is this supposed to be a joke?

satan !

this is cringe

Can you make your face appears larger on video? I mean, you're so cute, I even forgot what I'm searching here on yt

if you could've showed the sample format of notes.jsonl and notes_validateion.jsonl, it would've really helped, it' really confusing