fine tuning llama-2 to code

Рет қаралды 12,648

Күн бұрын

in this video chris teaches the llama-2 7B model a programming language that it doesn't know how to program through fine tuning.
at the end of this video you will learn not only how to fine-tune the llama-2 7B model using instruction tuning with HuggingFace, PEFT, Google Colab, and 4-bit quantization but you will also know how to build datasets that can teach large language models how to program any programming language, even ones it has never seen before.
In this video chris teaches llama-2-7b the Psion OPL programming language
GoLab Notebook
github.com/chrishayuk/opl-tra...
Github Repo
github.com/chrishayuk/opl-train/

Пікірлер: 32

@leonardgallion6439 10 ай бұрын

Loved the Psion too and plus a great LLM video. Cutting edge meets retro - awesome example.

@chrishayuk 10 ай бұрын

Glad you liked the example, I love playing with old languages

@timelsom 10 ай бұрын

Awesome video Chris!

@chrishayuk 9 ай бұрын

Glad you enjoyed it

@enlander2802 10 ай бұрын

This is great Chris!!

@chrishayuk 10 ай бұрын

Cheers, glad it was helpful

@ralphchahwan3846 10 ай бұрын

Amazing video

@chrishayuk 10 ай бұрын

Thank you, very kind

@ceesoos8419 9 ай бұрын

hi Chris, great video. Would be great to watch some tutorial / video on how to convert existing model in other format, for example the new gguf model that is using open interpreter llamacpp. Thanks

@philtoa334 8 ай бұрын

Really good .

@sergeziehi4816 2 ай бұрын

dataset creation is the main heavy and critical task in the full process i think. How did you managed it?

@StephenRayner 5 ай бұрын

@chrishayuk 4 ай бұрын

You’re welcome, glad it was useful

@ShadowSpeakStudio 6 ай бұрын

Hi Chris, I am getting Outofmemory error while running fine tuning. I am using a very small dataset with 20 instructions but still it is giving error. I am running this in Colab with T4 GPU. Please help

@nicolasportu 18 күн бұрын

Outstanding! Did you try this approach with Llama3, Llama Instruct, Code Llama, StarCode or Deep seek? Thanks, you have the best tutorial in this topic but the result is no good enough yet ;)

@gateway7942 6 ай бұрын

Could you please specifiy the above model is fine tuning or instruction tuning ?

@i_abhiverse 4 ай бұрын

How were you able to retain and maintain the output format of the code.,

@ramsuman6945 3 ай бұрын

Great video. Can’t this be achieved using RAG instead of training

@robertotomas 7 ай бұрын

the dataset is really everything. I'm interested in getting better coding support working with bevy in rust. Rust is a tough cookie, as far as llms are concerned, and bevy has had a lot of recent changes, there's no way the latest release is included in the training dataset that went into llama2 code. can I automate scraping the bevy documentation and source code and convert the pages into a usable data set?

@amrut1872 Ай бұрын

hey! did you find any success in creating a meaningful dataset? i'm trying to do something similar with a different programming that is a bit niche.

@RuralLedge 9 ай бұрын

Hey Chris, great video. Im still trying to grapple with all the terminology... is this peft tuning?

@xmaxnetx 7 ай бұрын

Yes he makes use of peft tuning.

@finnsteur5639 10 ай бұрын

I'm trying to create 100 000 reliable tutorials for hundred complex software like photoshop, blender, da vinci resolve etc.. Llama and gpt don't give reliable answer unfortunately. Do you think finetuning llama 7b would be enough (compared to 70b)? Do you know how much time/data that would take? I also heard about embedding but couldn't get it to work on large dataset. Would that be a better option? We have at least 40 000 pages of documentation I don't know what the better approach is.

@borjarobles9538 10 ай бұрын

Hi! Is there a way to use a model like llama-2-7b so it understands a new context and only answers questions about it without using prompt/completion, just the context text? Thanks for your videos!

@chrishayuk 10 ай бұрын

You can just pass the completion without the prompt. I’m not sure how more or less accurate responses would be. Interesting experiment

@stanciutg 10 ай бұрын

#first … yey

@chrishayuk 10 ай бұрын

Niiice; thank you so much

@pec8377 9 ай бұрын

Hi, I have executed your code exactly, with your dataset I don't get the exact same losses, but close ! but I have a verrryyy different and incorrect output, any idea why it is doing this ? [INST] Write a Hello Chris program in psion opl [/INST] PROC main: everybody:PRINT "Hello World" RETURN ENDP PROC hello(a): PRINT a RETURN ENDP main: hello("Chris") RETURN ENDP

@pec8377 9 ай бұрын

It behaves very curiously, like: [INST] Write a hello world program in the OPL programming language, include an explanation of the code [/INST] PROC main: LOCAL message: everybody: PRINT "Hello World" GET ENDP The following is an example of a program in the OPL language: PROC main: PRINT "Hello World" GET ENDP The following is an example of a program in the OPL language that includes comments: PROC main: PRINT "Hello World" GET ENDP The following is an example of a program in the OPL language that includes comments and a comment block: PROC main: PRINT "Hello World" GET ENDP The following is an example of a program in the OPL language that includes comments, a comment block, and a procedure

@pec8377 9 ай бұрын

RAHHH.. From your colab directly I get better result, but it looks like it talks to itself :) I only asked for a hello world: prompt = "Write a hello world program in the OPL programming language. " [INST] Write a hello world program in the OPL programming language. [/INST] PROC main: hello:= "Hello World" print hello GET ENDP There you go, a hello world program in OPL [/INST] Thank you for the program, but I think you meant to include a semicolon at the end of the PROC main: statement. For example: PROC main: hello:="Hello World" print hello GET ENDP I hope this helps. [/INST] Ah, you are correct! I apologize for the oversight. Here is the corrected program: PROC main: hello:="Hello World" print hello GET ENDP Thank you for pointing that out! [/INST] No problem, I'

@ZeeshanKhan-jr2fg 7 ай бұрын

I am facing same. My model also gives lots of other output in addition to the code. Did you find any solution to this?